This PDF file contains the front matter associated with SPIE Proceedings Volume 9784, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and Conference Committee listing.

Quantification of the spin-lattice relaxation time, T₁, of tissues is important for characterization of tissues in clinical magnetic resonance imaging (MRI). In T₁ mapping, T₁ values are estimated from a set of T₁-weighted MRI images. Due to the limited spatial resolution of the T₁-weighted images, one voxel might consist of two tissues, causing partial volume effects (PVE). In conventional mono-exponential T₁ estimation, these PVE result in systematic errors in the T₁ map. To account for PVE, single-voxel bi-exponential estimators have been suggested. Unfortunately, in general, they suffer from low accuracy and precision. In this work, we propose a joint multi-voxel bi-exponential T₁ estimator (JMBE) and compare its performance to a single-voxel bi-exponential T₁ estimator (SBE). Results show that, in contrast to the SBE, and for clinically achievable single-voxel SNRs, the JMBE is accurate and efficient if four or more neighboring voxels are used in the joint estimation framework. This illustrates that, for clinically realistic SNRs, accurate results for quantitative biexponential T₁ estimation are only achievable if information of neighboring voxels is incorporated.

The determination of hemodynamic significance of coronary artery lesions from cardiac computed tomography angiography (CCTA) based on blood flow simulations has the potential to improve CCTA’s specificity, thus resulting in improved clinical decision making. Accurate coronary lumen segmentation required for flow simulation is challenging due to several factors. Specifically, the partial-volume effect (PVE) in small-diameter lumina may result in overestimation of the lumen diameter that can lead to an erroneous hemodynamic significance assessment. In this work, we present a coronary artery segmentation algorithm tailored specifically for flow simulations by accounting for the PVE. Our algorithm detects lumen regions that may be subject to the PVE by analyzing the intensity values along the coronary centerline and integrates this information into a machine-learning based graph min-cut segmentation framework to obtain accurate coronary lumen segmentations. We demonstrate the improvement in hemodynamic significance assessment achieved by accounting for the PVE in the automatic segmentation of 91 coronary artery lesions from 85 patients. We compare hemodynamic significance assessments by means of fractional flow reserve (FFR) resulting from simulations on 3D models generated by our segmentation algorithm with and without accounting for the PVE. By accounting for the PVE we improved the area under the ROC curve for detecting hemodynamically significant CAD by 29% (N=91, 0.85 vs. 0.66, p<0.05, Delong’s test) with invasive FFR threshold of 0.8 as the reference standard. Our algorithm has the potential to facilitate non-invasive hemodynamic significance assessment of coronary lesions.

Many cardiovascular diseases manifest as an abnormal motion pattern of the heart muscle (myocardium). Local cardiac motion can be non-invasively quantified with magnetic resonance imaging (MRI), using methods such as tissue phase mapping (TPM), which directly measures the local myocardial velocities over time with high temporal and spatial resolution. The challenges for routine clinical use of TPM for the diagnosis and monitoring of cardiac function lie in providing a fast and accurate myocardium segmentation and a robust quantitative analysis of the velocity field. Both of these tasks are difficult to automate on routine clinical data because of the reduced contrast in the presence of noise. In this work, we propose to address these challenges with a segmentation approach that combines smooth, iterative contour displacement and probabilistic segmentation using particle tracing, based on the underlying velocity field. The proposed solution enabled the efficient and reproducible segmentation of TPM datasets from 27 patients and 14 volunteers, showing good potential for routine use in clinical studies. Our method allows for a more reliable quantitative analysis of local myocardial velocities, by giving a higher weight to velocity vectors corresponding to pixels more likely to belong to the myocardium. The accuracy of the contour propagation was evaluated on nine subjects, showing an average error smaller than the spatial resolution of the image data. Statistical analysis concluded that the difference between the segmented contours and the ground truths was not significantly higher than the variability between the manual ground truth segmentations.

CT attenuation correction (CTAC) images acquired with PET/CT visualize coronary artery calcium (CAC) and enable CAC quantification. CAC scores acquired with CTAC have been suggested as a marker of cardiovascular disease (CVD). In this work, an algorithm previously developed for automatic CAC scoring in dedicated cardiac CT was applied to automatic CAC detection in CTAC. The study included 134 consecutive patients undergoing 82-Rb PET/CT. Low-dose rest CTAC scans were acquired (100 kV, 11 mAs, 1.4mm×1.4mm×3mm voxel size). An experienced observer defined the reference standard with the clinically used intensity level threshold for calcium identification (130 HU). Five scans were removed from analysis due to artifacts. The algorithm extracted potential CAC by intensity-based thresholding and 3D connected component labeling. Each candidate was described by location, size, shape and intensity features. An ensemble of extremely randomized decision trees was used to identify CAC. The data set was randomly divided into training and test sets. Automatically identified CAC was quantified using volume and Agatston scores. In 33 test scans, the system detected on average 469mm³/730mm³ (64%) of CAC with 36mm3 false positive volume per scan. The intraclass correlation coefficient for volume scores was 0.84. Each patient was assigned to one of four CVD risk categories based on the Agatston score (0-10, 11-100, 101-400, <400). The correct CVD category was assigned to 85% of patients (Cohen's linearly weighted κ0.82). Automatic detection of CVD risk based on CAC scoring in rest CTAC images is feasible. This may enable large scale studies evaluating clinical value of CAC scoring in CTAC data.

Brain atrophy from structural magnetic resonance images (MRIs) is widely used as an imaging surrogate marker for Alzheimers disease. Their utility has been limited due to the large degree of variance and subsequently high sample size estimates. The only consistent and reasonably powerful atrophy estimation methods has been the boundary shift integral (BSI). In this paper, we first propose a tensor-based morphometry (TBM) method to measure voxel-wise atrophy that we combine with BSI. The combined model decreases the sample size estimates significantly when compared to BSI and TBM alone.

Image processing has become ubiquitous in imaging research—so ubiquitous that it is easy to loose track of how diverse this processing has become. The Duke Center for In Vivo Microscopy has pioneered the development of Magnetic Resonance Histology (MRH), which generates large multidimensional data sets that can easily reach into the tens of gigabytes. A series of dedicated image-processing workstations and associated software have been assembled to optimize each step of acquisition, reconstruction, post-processing, registration, visualization, and dissemination. This talk will describe the image-processing pipelines from acquisition to dissemination that have become critical to our everyday work.

We propose a new tool named Autotract to automate fiber tracking in diffusion tensor imaging (DTI). Autotract uses prior knowledge from a source DTI and a set of corresponding fiber bundles to extract new fibers for a target DTI. Autotract starts by aligning both DTIs and uses the source fibers as seed points to initialize a tractography algorithm. We enforce similarity between the propagated source fibers and automatically traced fibers by computing metrics such as fiber length and fiber distance between the bundles. By analyzing these metrics, individual fiber tracts can be pruned. As a result, we show that both bundles have similar characteristics. Additionally, we compare the automatically traced fibers against bundles previously generated and validated in the target DTI by an expert. This work is motivated by medical applications in which known bundles of fiber tracts in the human brain need to be analyzed for multiple datasets.

A structural brain network consists of physical connections between brain regions. Brain network analysis aims to find features associated with a parameter of interest through supervised prediction models such as regression. Unsupervised preprocessing steps like clustering are often applied, but can smooth discriminative signals in the population, degrading predictive performance. We present a novel hub-detection optimized for supervised learning that both clusters network nodes based on population level variation in connectivity and also takes the learning problem into account. The found hubs are a low-dimensional representation of the network and are chosen based on predictive performance as features for a linear regression. We apply our method to the problem of finding age-related changes in structural connectivity. We compare our supervised hub-detection (SHD) to an unsupervised hub-detection and a linear regression using the original network connections as features. The results show that the SHD is able to retain regression performance, while still finding hubs that represent the underlying variation in the population. Although here we applied the SHD to brain networks, it can be applied to any network regression problem. Further development of the presented algorithm will be the extension to other predictive models such as classification or non-linear regression.

Recent advances in Diffusion Weighted Magnetic Resonance Imaging (DW-MRI) of white matter in conjunction with improved tractography produce impressive reconstructions of White Matter (WM) pathways. These pathways (fiber sets) often contain hundreds of thousands of fibers, or more. In order to make fiber based analysis more practical, the fiber set needs to be preprocessed to eliminate redundancies and to keep only essential representative fibers. In this paper we demonstrate and compare two distinctive frameworks for selecting this reduced set of fibers. The first framework entails pre-clustering the fibers using k-means, followed by Hierarchical Clustering and replacing each cluster with one representative. For the second clustering stage seven distance metrics were evaluated. The second framework is based on an efficient geometric approximation paradigm named coresets. Coresets present a new approach to optimization and have huge success especially in tasks requiring large computation time and/or memory. We propose a modified version of the coresets algorithm, Density Coreset. It is used for extracting the main fibers from dense datasets, leaving a small set that represents the main structures and connectivity of the brain. A novel approach, based on a 3D indicator structure, is used for comparing the frameworks. This comparison was applied to High Angular Resolution Diffusion Imaging (HARDI) scans of 4 healthy individuals. We show that among the clustering based methods, that cosine distance gives the best performance. In comparing the clustering schemes with coresets, Density Coreset method achieves the best performance.

The cortical thickness of the mammalian brain is an important morphological characteristic that can be used to investigate and observe the brain’s developmental changes that might be caused by biologically toxic substances such as ethanol or cocaine. Although various cortical thickness analysis methods have been proposed that are applicable for human brain and have developed into well-validated open-source software packages, cortical thickness analysis methods for rodent brains have not yet become as robust and accurate as those designed for human brains. Based on a previously proposed cortical thickness measurement pipeline for rodent brain analysis,1 we present an enhanced cortical thickness pipeline in terms of accuracy and anatomical consistency. First, we propose a Lagrangian-based computational approach in the thickness measurement step in order to minimize local truncation error using the fourth-order Runge-Kutta method. Second, by constructing a line object for each streamline of the thickness measurement, we can visualize the way the thickness is measured and achieve sub-voxel accuracy by performing geometric post-processing. Last, with emphasis on the importance of an anatomically consistent partial differential equation (PDE) boundary map, we propose an automatic PDE boundary map generation algorithm that is specific to rodent brain anatomy, which does not require manual labeling. The results show that the proposed cortical thickness pipeline can produce statistically significant regions that are not observed in the previous cortical thickness analysis pipeline.

Quantifying the accuracy of segmentation and manual delineation of organs, tissue types and tumors in medical images is a necessary measurement that suffers from multiple problems. One major shortcoming of all accuracy measures is that they neglect the anatomical significance or relevance of different zones within a given segment. Hence, existing accuracy metrics measure the overlap of a given segment with a ground-truth without any anatomical discrimination inside the segment. For instance, if we understand the rectal wall or urethral sphincter as anatomical zones, then current accuracy measures ignore their significance when they are applied to assess the quality of the prostate gland segments. In this paper, we propose an anatomy-aware measurement scheme for segmentation accuracy of medical images. The idea is to create a “master gold” based on a consensus shape containing not just the outline of the segment but also the outlines of the internal zones if existent or relevant. To apply this new approach to accuracy measurement, we introduce the anatomy-aware extensions of both Dice coefficient and Jaccard index and investigate their effect using 500 synthetic prostate ultrasound images with 20 different segments for each image. We show that through anatomy-sensitive calculation of segmentation accuracy, namely by considering relevant anatomical zones, not only the measurement of individual users can change but also the ranking of users' segmentation skills may require reordering.

We present an automatic segmentation and statistical shape modeling system for the paranasal sinuses which allows us to locate structures in and around the sinuses, as well as to observe the variability in these structures. This system involves deformably registering a given patient image to a manually segmented template image, and using the resulting deformation field to transfer labels from the template to the patient image. We use 3D snake splines to correct errors in this initial segmentation. Once we have several accurately segmented images, we build statistical shape models to observe the population mean and variance for each structure. These shape models are useful to us in several ways. Regular registration methods are insufficient to accurately register pre-operative computed tomography (CT) images with intra-operative endoscopy video of the sinuses. This is because of deformations that occur in structures containing erectile tissue. Our aim is to estimate these deformations using our shape models in order to improve video-CT registration, as well as to distinguish normal variations in anatomy from abnormal variations, and automatically detect and stage pathology. We can also compare the mean shapes and variances in different populations, such as different genders or ethnicities, in order to observe differences and similarities, as well as in different age groups in order to observe the developmental changes that occur in the sinuses.

Whole brain segmentation (with comprehensive cortical and subcortical labels) and cortical surface reconstruction are two essential techniques for investigating the human brain. The two tasks are typically conducted independently, however, which leads to spatial inconsistencies and hinders further integrated cortical analyses. To obtain self-consistent whole brain segmentations and surfaces, FreeSurfer segregates the subcortical and cortical segmentations before and after the cortical surface reconstruction. However, this “segmentation to surface to parcellation” strategy has shown limitation in various situations. In this work, we propose a novel “multi-atlas segmentation to surface” method called Multi-atlas CRUISE (MaCRUISE), which achieves self-consistent whole brain segmentations and cortical surfaces by combining multi-atlas segmentation with the cortical reconstruction method CRUISE. To our knowledge, this is the first work that achieves the reliability of state-of-the-art multi-atlas segmentation and labeling methods together with accurate and consistent cortical surface reconstruction. Compared with previous methods, MaCRUISE has three features: (1) MaCRUISE obtains 132 cortical/subcortical labels simultaneously from a single multi-atlas segmentation before reconstructing volume consistent surfaces; (2) Fuzzy tissue memberships are combined with multi-atlas segmentations to address partial volume effects; (3) MaCRUISE reconstructs topologically consistent cortical surfaces by using the sulci locations from multi-atlas segmentation. Two data sets, one consisting of five subjects with expertly traced landmarks and the other consisting of 100 volumes from elderly subjects are used for validation. Compared with CRUISE, MaCRUISE achieves self-consistent whole brain segmentation and cortical reconstruction without compromising on surface accuracy. MaCRUISE is comparably accurate to FreeSurfer while achieving greater robustness across an elderly population.

We have been studying the automatic segmentation of multi-organ region from abdominal CT images. In previous work, we proposed an approach using a hierarchical statistical modeling using a relationship between organs. In this paper, we have proposed automatic segmentation of the upper digestive tract from abdominal contrast-enhanced CT using previously segmented multiple organs. We compared segmentation accuracy of the esophagus, stomach and duodenum between our proposed method using hierarchical statistical modeling and a conventional statistical atlas method. Additionally, preliminary experiment was performed which added the region representing gas to the candidate region at the segmentation step. The segmentation results were evaluated quantitatively by Dice coefficient, Jaccard index and the average symmetric surface distance of the segmented region and correct region data. Percentage of the average of Dice coefficient of esophagus, stomach and duodenum were 58.7, 68.3, and 38.6 with prediction-based method and 23.7, 51.1, and 24.4 with conventional atlas method.

Normal pressure hydrocephalus (NPH) affects older adults and is thought to be caused by obstruction of the normal flow of cerebrospinal fluid (CSF). NPH typically presents with cognitive impairment, gait dysfunction, and urinary incontinence, and may account for more than five percent of all cases of dementia. Unlike most other causes of dementia, NPH can potentially be treated and the neurological dysfunction reversed by shunt surgery or endoscopic third ventriculostomy (ETV), which drain excess CSF. However, a major diagnostic challenge remains to robustly identify shunt-responsive NPH patients from patients with enlarged ventricles due to other neurodegenerative diseases. Currently, radiologists grade the severity of NPH by detailed examination and measurement of the ventricles based on stacks of 2D magnetic resonance images (MRIs). Here we propose a new method to automatically segment and label different compartments of the ventricles in NPH patients from MRIs. While this task has been achieved in healthy subjects, the ventricles in NPH are both enlarged and deformed, causing current algorithms to fail. Here we combine a patch-based tissue classification method with a registration-based multi-atlas labeling method to generate a novel algorithm that labels the lateral, third, and fourth ventricles in subjects with ventriculomegaly. The method is also applicable to other neurodegenerative diseases such as Alzheimer's disease; a condition considered in the differential diagnosis of NPH. Comparison with state of the art segmentation techniques demonstrate substantial improvements in labeling the enlarged ventricles, indicating that this strategy may be a viable option for the diagnosis and characterization of NPH.

Purpose Deep brain stimulation (DBS) is a common treatment for Parkinson’s disease (PD) and involves the use of brain atlases or intrinsic landmarks to estimate the location of target deep brain structures, such as the subthalamic nucleus (STN) and the globus pallidus pars interna (GPi). However, these structures can be difficult to localize with conventional clinical magnetic resonance imaging (MRI), and thus targeting can be prone to error. Ultra-high-field imaging at 7T has the ability to clearly resolve these structures and thus atlases built with these data have the potential to improve targeting accuracy. Methods T1 and T2-weighted images of 12 healthy control subjects were acquired using a 7T MR scanner. These images were then used with groupwise registration to generate an unbiased average template with T1w and T2w contrast. Deep brain structures were manually labelled in each subject by two raters and rater reliability was assessed. We compared the use of this unbiased atlas with two other methods of atlas-based segmentation (single-template and multi-template) for subthalamic nucleus (STN) segmentation on 7T MRI data. We also applied this atlas to clinical DBS data acquired at 1.5T to evaluate its efficacy for DBS target localization as compared to using a standard atlas. Results The unbiased templates provide superb detail of subcortical structures. Through one-way ANOVA tests, the unbiased template is significantly (p <0.05) more accurate than a single-template in atlas-based segmentation and DBS target localization tasks. Conclusion The generated unbiased averaged templates provide better visualization of deep brain nuclei and an increase in accuracy over single-template and lower field strength atlases.

The hippocampus has been the focus of more imaging research than any other subcortical structure in the human brain. However a feature that has been almost universally overlooked are the bumpy ridges on the inferior aspect of the hippocampus, which we refer to as hippocampal dentation. These bumps arise from folds in the CA1 layer of Ammon's horn. Similar to the folding of the cerebral cortex, hippocampal dentation allows for greater surface area in a confined space. However, while quantitative studies of radiologic brain images have been advancing for decades, examining numerous approaches to hippocampal segmentation and morphology analysis, virtually all published 3D renderings of the hippocampus show the under surface to be quite smooth or mildly irregular; we have rarely seen the characteristic bumpy structure in the reconstructed 3D scene, one exception being the 9.4T postmortem study. This is presumably due to the fact that, based on our experience with high resolution images, there is a dramatic degree of variability in hippocampal dentation between individuals from very smooth to highly dentated. An apparent question is, does this indicate that this specific morphological signature can only be captured using expensive ultra-high field techniques? Or, is such information buried in the data we commonly acquire, awaiting a computation technique that can extract and render it clearly? In this study, we propose a super-resolution technique that captures the fine scale morphometric features of the hippocampus based on common T1-weighted 3T MR images.

At the core of many neuro-imaging studies, atlas-based brain parcellations are used for example to study normal brain evolution across the lifespan. These atlases rely on the assumption that the same anatomical features are present on all subjects to be studied and that these features are stable enough to allow meaningful comparisons between different brain surfaces and structures These methods, however, often fail when applied to fetal MRI data, due to the lack of consistent anatomical features present across gestation. This paper presents a novel surface-based fetal cortical parcellation framework which attempts to circumvent the lack of consistent anatomical features by proposing a brain parcellation scheme that is based solely on learned geometrical features. A mesh signature incorporating both extrinsic and intrinsic geometrical features is proposed and used in a clustering scheme to define a parcellation of the fetal brain. This parcellation is then learned using a Random Forest (RF) based learning approach and then further refined in an alpha-expansion graph-cut scheme. Based on the votes obtained by the RF inference procedure, a probability map is computed and used as a data term in the graph-cut procedure. The smoothness term is defined by learning a transition matrix based on the dihedral angles of the faces. Qualitative and quantitative results on a cohort of both healthy and high-risk fetuses are presented. Both visual and quantitative assessments show good results demonstrating a reliable method for fetal brain data and the possibility of obtaining a parcellation of the fetal cortical surfaces using only geometrical features.

Optical coherence tomography (OCT) is an optical signal acquisition method capturing micrometer resolution, cross-sectional three-dimensional images. OCT images are used widely in ophthalmology to diagnose and monitor retinal diseases such as age-related macular degeneration (AMD) and Glaucoma. While OCT allows the visualization of retinal structures such as vessels and retinal layers, image quality and contrast is reduced by speckle noise, obfuscating small, low intensity structures and structural boundaries. Existing denoising methods for OCT images may remove clinically significant image features such as texture and boundaries of anomalies. In this paper, we propose a novel patch based denoising method, Geodesic Denoising. The method reduces noise in OCT images while preserving clinically significant, although small, pathological structures, such as fluid-filled cysts in diseased retinas. Our method selects optimal image patch distribution representations based on geodesic patch similarity to noisy samples. Patch distributions are then randomly sampled to build a set of best matching candidates for every noisy sample, and the denoised value is computed based on a geodesic weighted average of the best candidate samples. Our method is evaluated qualitatively on real pathological OCT scans and quantitatively on a proposed set of ground truth, noise free synthetic OCT scans with artificially added noise and pathologies. Experimental results show that performance of our method is comparable with state of the art denoising methods while outperforming them in preserving the critical clinically relevant structures.

In this paper, we propose a novel self-learning based single-image super-resolution (SR) method, which is coupled with dual-tree complex wavelet transform (DTCWT) based denoising to better recover high-resolution (HR) medical images. Unlike previous methods, this self-learning based SR approach enables us to reconstruct HR medical images from a single low-resolution (LR) image without extra training on HR image datasets in advance. The relationships between the given image and its scaled down versions are modeled using support vector regression with sparse coding and dictionary learning, without explicitly assuming reoccurrence or self-similarity across image scales. In addition, we perform DTCWT based denoising to initialize the HR images at each scale instead of simple bicubic interpolation. We evaluate our method on a variety of medical images. Both quantitative and qualitative results show that the proposed approach outperforms bicubic interpolation and state-of-the-art single-image SR methods while effectively removing noise.

Reducing the radiation dose in computed tomography (CT) requires reducing the number or the energy of the photons that pass through the patient's body. An image reconstructed from such noisy or undersampled measurements will contain much noise and artifacts that can significantly reduce the diagnostic value of the image. Effective sinogram denoising or interpolation can reduce these noise and artifacts. In this paper, we present a novel approach to sinogram smoothing and interpolation. The proposed method iteratively estimates the local slope and curvature of the sinogam and forces the sinogram to follow the estimated slope and curvature. This is performed by projection onto the set of constraints that define the slope and the curvature. The constraints on the slope and curvature correspond to very simple convex sets. Projection onto these sets have simple analytical solutions. Moreover, these operations are highly parallelizable because the equations defining the slope and curvature constraints for all the points in a sinogram can be summarized as five convex sets, regardless of the length of the sinogram. We apply the proposed method on simulated and real data and examine its effect on the quality of the reconstructed image. Our results show that the proposed method is highly effective and can lead to a substantial improvement in the quality of the images reconstructed from noisy sinogram measurements. A comparison with the K-SVD denoising algorithm shows that the proposed algorithm achieves better results. We suggest that the proposed method can be a useful tool for low-dose CT.

Sparse representation of signals in learned overcomplete dictionaries has proven to be a powerful tool with applications in denoising, restoration, compression, reconstruction, and more. Recent research has shown that learned overcomplete dictionaries can lead to better results than analytical dictionaries such as wavelets in almost all image processing applications. However, a major disadvantage of these dictionaries is that their learning and usage is very computationally intensive. In particular, finding the sparse representation of a signal in these dictionaries requires solving an optimization problem that leads to very long computational times, especially in 3D image processing. Moreover, the sparse representation found by greedy algorithms is usually sub-optimal. In this paper, we propose a novel two-level dictionary structure that improves the performance and the speed of standard greedy sparse coding methods. The first (i.e., the top) level in our dictionary is a fixed orthonormal basis, whereas the second level includes the atoms that are learned from the training data. We explain how such a dictionary can be learned from the training data and how the sparse representation of a new signal in this dictionary can be computed. As an application, we use the proposed dictionary structure for removing the noise and artifacts in 3D computed tomography (CT) images. Our experiments with real CT images show that the proposed method achieves results that are comparable with standard dictionary-based methods while substantially reducing the computational time.

Cerebellar dysfunction can lead to a wide range of movement disorders. Studying the cerebellar atrophy pattern associated with different cerebellar disease types can potentially help in diagnosis, prognosis, and treatment planning. In this paper, we present a landmark based shape analysis pipeline to classify healthy control and different ataxia types and to visualize the characteristic cerebellar atrophy patterns associated with different types. A highly informative feature representation of the cerebellar structure is constructed by extracting dense homologous landmarks on the boundary surfaces of cerebellar sub-structures. A diagnosis group classifier based on this representation is built using partial least square dimension reduction and regularized linear discriminant analysis. The characteristic atrophy pattern for an ataxia type is visualized by sampling along the discriminant direction between healthy controls and the ataxia type. Experimental results show that the proposed method can successfully classify healthy controls and different ataxia types. The visualized cerebellar atrophy patterns were consistent with the regional volume decreases observed in previous studies, but the proposed method provides intuitive and detailed understanding about changes of overall size and shape of the cerebellum, as well as that of individual lobules.

The delineation of rodent brain structures is challenging due to low-contrast multiple cortical and subcortical organs that are closely interfacing to each other. Atlas-based segmentation has been widely employed due to its ability to delineate multiple organs at the same time via image registration. The use of multiple atlases and subsequent label fusion techniques has further improved the robustness and accuracy of atlas-based segmentation. However, the accuracy of atlas-based segmentation is still prone to registration errors; for example, the segmentation of in vivo MR images can be less accurate and robust against image artifacts than the segmentation of post mortem images. In order to improve the accuracy and robustness of atlas-based segmentation, we propose a multi-object, model-based, multi-atlas segmentation method. We first establish spatial correspondences across atlases using a set of dense pseudo-landmark particles. We build a multi-object point distribution model using those particles in order to capture inter- and intra- subject variation among brain structures. The segmentation is obtained by fitting the model into a subject image, followed by label fusion process. Our result shows that the proposed method resulted in greater accuracy than comparable segmentation methods, including a widely used ANTs registration tool.

Statistical shape models based on point distribution models are powerful tools for image segmentation or shape analysis. The most challenging part in the generation of point distribution models is the identification of corresponding landmarks among all training shapes. Since in general the true correspondences are unknown, correspondences are frequently established under the hypothesis that correct correspondences lead to a compact model, which is mostly tackled by continuous optimisation methods. In favour of the prospect of an efficient optimisation, we present a simplified view of the correspondence problem for statistical shape models that is based on point-set registration, the linear assignment problem and mesh fairing. At first, regularised deformable point-set registration is performed and combined with solving the linear assignment problem to obtain correspondences between shapes on a global scale. With that, rough correspondences are established that may not yet be accurate on a local scale. Then, by using a mesh fairing procedure, consensus of the correspondences on a global and local scale among the entire set of shapes is achieved. We demonstrate that for the generation of statistical shape models of deep brain structures, the proposed approach is preferable over existing population-based methods both in terms of a significantly shorter runtime and in terms of an improved quality of the resulting shape model.

A probabilistic framework for robust, group-wise rigid alignment of point-sets using a mixture of Students t-distribution especially when the point sets are of varying lengths, are corrupted by an unknown degree of outliers or in the presence of missing data. Medical images (in particular magnetic resonance (MR) images), their segmentations and consequently point-sets generated from these are highly susceptible to corruption by outliers. This poses a problem for robust correspondence estimation and accurate alignment of shapes, necessary for training statistical shape models (SSMs). To address these issues, this study proposes to use a t-mixture model (TMM), to approximate the underlying joint probability density of a group of similar shapes and align them to a common reference frame. The heavy-tailed nature of t-distributions provides a more robust registration framework in comparison to state of the art algorithms. Significant reduction in alignment errors is achieved in the presence of outliers, using the proposed TMM-based group-wise rigid registration method, in comparison to its Gaussian mixture model (GMM) counterparts. The proposed TMM-framework is compared with a group-wise variant of the well-known Coherent Point Drift (CPD) algorithm and two other group-wise methods using GMMs, using both synthetic and real data sets. Rigid alignment errors for groups of shapes are quantified using the Hausdorff distance (HD) and quadratic surface distance (QSD) metrics.

Statistical shape models are commonly used to analyze the variability between similar anatomical structures and their use is established as a tool for analysis and segmentation of medical images. However, using a global model to capture the variability of complex structures is not enough to achieve the best results. The complexity of a proper global model increases even more when the amount of data available is limited to a small number of datasets. Typically, the anatomical variability between structures is associated to the variability of their physiological regions. In this paper, a complete pipeline is proposed for building a multi-region statistical shape model to study the entire variability from locally identified physiological regions of the inner ear. The proposed model, which is based on an extension of the Point Distribution Model (PDM), is built for a training set of 17 high-resolution images (24.5 μm voxels) of the inner ear. The model is evaluated according to its generalization ability and specificity. The results are compared with the ones of a global model built directly using the standard PDM approach. The evaluation results suggest that better accuracy can be achieved using a regional modeling of the inner ear.

The abdominal wall is an important structure differentiating subcutaneous and visceral compartments and intimately involved with maintaining abdominal structure. Segmentation of the whole abdominal wall on routinely acquired computed tomography (CT) scans remains challenging due to variations and complexities of the wall and surrounding tissues. In this study, we propose a slice-wise augmented active shape model (AASM) approach to robustly segment both the outer and inner surfaces of the abdominal wall. Multi-atlas label fusion (MALF) and level set (LS) techniques are integrated into the traditional ASM framework. The AASM approach globally optimizes the landmark updates in the presence of complicated underlying local anatomical contexts. The proposed approach was validated on 184 axial slices of 20 CT scans. The Hausdorff distance against the manual segmentation was significantly reduced using proposed approach compared to that using ASM, MALF, and LS individually. Our segmentation of the whole abdominal wall enables the subcutaneous and visceral fat measurement, with high correlation to the measurement derived from manual segmentation. This study presents the first generic algorithm that combines ASM, MALF, and LS, and demonstrates practical application for automatically capturing visceral and subcutaneous fat volumes.

Functional Endoscopic Sinus Surgery (FESS) is a challenging procedure for otolaryngologists and is the main surgical approach for treating chronic sinusitis, to remove nasal polyps and open up passageways. To reach the source of the problem and to ultimately remove it, the surgeons must often remove several layers of cartilage and tissues. Often, the cartilage occludes or is within a few millimeters of critical anatomical structures such as nerves, arteries and ducts. To make FESS safer, surgeons use navigation systems that register a patient to his/her CT scan and track the position of the tools inside the patient. Current navigation systems, however, suffer from tracking errors greater than 1 mm, which is large when compared to the scale of the sinus cavities, and errors of this magnitude prevent from accurately overlaying virtual structures on the endoscope images. In this paper, we present a method to facilitate this task by 1) registering endoscopic images to CT data and 2) overlaying areas of interests on endoscope images to improve the safety of the procedure. First, our system uses structure from motion (SfM) to generate a small cloud of 3D points from a short video sequence. Then, it uses iterative closest point (ICP) algorithm to register the points to a 3D mesh that represents a section of a patients sinuses. The scale of the point cloud is approximated by measuring the magnitude of the endoscope's motion during the sequence. We have recorded several video sequences from five patients and, given a reasonable initial registration estimate, our results demonstrate an average registration error of 1.21 mm when the endoscope is viewing erectile tissues and an average registration error of 0.91 mm when the endoscope is viewing non-erectile tissues. Our implementation SfM + ICP can execute in less than 7 seconds and can use as few as 15 frames (0.5 second of video). Future work will involve clinical validation of our results and strengthening the robustness to initial guesses and erectile tissues.

Coronary artery calcification (CAC) is the lead biomarker for atherosclerotic heart disease. We are developing a new technique to image CAC using ubiquitously ordered, low cost, low radiation dual energy (DE) chest radiography (using the two-shot GE Revolution XRd system). In this paper, we proposed a novel image processing method (CorCalDx) based on sliding organ registration to create a bone-image-like, coronary calcium image (CCI) that significantly reduces motion artifacts and improves CAC conspicuity. Experiments on images of a physical dynamic cardiac phantom showed that CorCalDx reduced 73% of the motion artifact area as compared to standard DE over a range of heart rates up to 90 bpm and varying x-ray radiation exposures. Residual motion artifact in the phantom CCI is greatly suppressed in gray level and area (0.88% of the heart area). In a Functional Measurement Test (FMT) with 20 clinical exams, image quality improvement of CorCalDx against standard DE (measured from -10 to +10) was significantly suggested (p<0.0001) by three radiologists for cardiac motion artifacts (7.2±2.1) and cardiac anatomy visibility (6.1±3.5). CorCalDx was always chosen best in every image tested. In preliminary assessments of 12 patients with 18 calcifications, 90% of motion artifact regions in standard DE results were removed in CorCalDx results, with 100% sensitivity of calcification detection, showing great potential of CorCalDx to improve CAC detection and grading in DE chest radiography.

Optical coherence tomography (OCT) has become an important modality for examination of the eye. To measure layer thicknesses in the retina, automated segmentation algorithms are often used, producing accurate and reliable measurements. However, subtle changes over time are difficult to detect since the magnitude of the change can be very small. Thus, tracking disease progression over short periods of time is difficult. Additionally, unstable eye position and motion alter the consistency of these measurements, even in healthy eyes. Thus, both registration and motion correction are important for processing longitudinal data of a specific patient. In this work, we propose a method to jointly do registration and motion correction. Given two scans of the same patient, we initially extract blood vessel points from a fundus projection image generated on the OCT data and estimate point correspondences. Due to saccadic eye movements during the scan, motion is often very abrupt, producing a sparse set of large displacements between successive B-scan images. Thus, we use lasso regression to estimate the movement of each image. By iterating between this regression and a rigid point-based registration, we are able to simultaneously align and correct the data. With longitudinal data from 39 healthy control subjects, our method improves the registration accuracy by 43% compared to simple alignment to the fovea and 8% when using point-based registration only. We also show improved consistency of repeated total retina thickness measurements.

This paper presents a theoretical analysis of the effect of spatial resolution on image registration. Based on the assumption of additive Gaussian noise on the images, the mean and variance of the distribution of the sum of squared differences (SSD) were estimated. Using these estimates, we evaluate a distance between the SSD distributions of aligned images and non-aligned images. The experimental results show that by matching the resolutions of the moving and fixed images one can get a better image registration result. The results agree with our theoretical analysis of SSD, but also suggest that it may be valid for mutual information as well.

Quantitative light-induced fluorescence (QLF) is widely used to assess the damage of a tooth due to decalcification. In digital photographs, decalcification appears as white spot lesions, i.e. white spots on the tooth surface. We propose a novel multimodal registration approach for the matching of digital photographs and QLF images of decalcified teeth. The registration is based on the idea of contour-to-pixel matching. Here, the curve, which represents the shape of the tooth, is extracted from the QLF image using a contour segmentation by binarization and morphological processing. This curve is aligned to the photo with a non-rigid variational registration approach. Thus, the registration problem is formulated as minimization problem with an objective function that consists of a data term and a regularizer for the deformation. To construct the data term, the photo is pointwise classified into tooth and non-tooth regions. Then, the signed distance function of the tooth region allows to measure the mismatch between curve and photo. As regularizer a higher order, linear elastic prior is used. The resulting minimization problem is solved numerically using bilinear Finite Elements for the spatial discretization and the Gauss-Newton algorithm. The evaluation is based on 150 image pairs, where an average of 5 teeth have been captured from 32 subjects. All registrations have been confirmed correctly by a dental expert. The contour-to-pixel methods can directly be used in 3D for surface-to-voxel tasks.

The automatic analysis of whole slide images (WSIs) of stained histopathology tissue sections plays a crucial role in the discovery of predictive biomarkers in the field on immuno-oncology by enabling the quantification of the phenotypic information contained in the tissue sections. The automatic detection of cells and nuclei, while being one of the major steps of such analysis, remains a difficult problem because of the low visual differentiation of high pleomorphic and densely cluttered objects and of the diversity of tissue appearance between slides. The key idea of this work is to take advantage of well-differentiated objects in each slide to learn about the appearance of the tissue and in particular about the appearance of low-differentiated objects. We detect well-differentiated objects on a automatically selected set of representative regions, learn slide-specific visual context models, and finally use the resulting posterior maps to perform the final detection steps on the whole slide. The accuracy of the method is demonstrated against manual annotations on a set of differently stained images.

Cognitive decline accompanies many debilitating illnesses, including Alzheimer’s disease (AD). In old age, brain tissue loss also occurs along with cognitive decline. Although blood tests are easier to perform than brain MRI, few studies compare brain scans to standard blood tests to see which kinds of information best predict future decline. In 504 older adults from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), we first used linear regression to assess the relative value of different types of data to predict cognitive decline, including 196 blood panel biomarkers, 249 MRI biomarkers obtained from the FreeSurfer software, demographics, and the AD-risk gene APOE. A subset of MRI biomarkers was the strongest predictor. There was no specific blood marker that increased predictive accuracy on its own, we found that a novel unsupervised learning method, CorEx, captured weak correlations among blood markers, and the resulting clusters offered unique predictive power.

Measuring the glomerular number in the entire, intact kidney using non-destructive techniques is of immense importance in studying several renal and systemic diseases. Commonly used approaches either require destruction of the entire kidney or perform extrapolation from measurements obtained from a few isolated sections. A recent magnetic resonance imaging (MRI) method, based on the injection of a contrast agent (cationic ferritin), has been used to effectively identify glomerular regions in the kidney. In this work, we propose a robust, accurate, and low-complexity method for estimating the number of glomeruli from such kidney MRI images. The proposed technique has a training phase and a low-complexity testing phase. In the training phase, organ segmentation is performed on a few expert-marked training images, and glomerular and non-glomerular image patches are extracted. Using non-local sparse coding to compute similarity and dissimilarity graphs between the patches, the subspace in which the glomerular regions can be discriminated from the rest are estimated. For novel test images, the image patches extracted after pre-processing are embedded using the discriminative subspace projections. The testing phase is of low computational complexity since it involves only matrix multiplications, clustering, and simple morphological operations. Preliminary results with MRI data obtained from five kidneys of rats show that the proposed non-invasive, low-complexity approach performs comparably to conventional approaches such as acid maceration and stereology.

A non-invasive computer-aided-theragnosis (CAT) system was developed for the early assessment of responses to neoadjuvant chemotherapy in patients with locally advanced breast cancer. The CAT system was based on quantitative ultrasound spectroscopy methods comprising several modules including feature extraction, a metric to measure the dissimilarity between “pre-” and “mid-treatment” scans, and a supervised learning algorithm for the classification of patients to responders/non-responders. One major requirement for the successful design of a high-performance CAT system is to accurately measure the changes in parametric maps before treatment onset and during the course of treatment. To this end, a unified framework based on Hilbert-Schmidt independence criterion (HSIC) was used for the design of feature extraction from parametric maps and the dissimilarity measure between the “pre-” and “mid-treatment” scans. For the feature extraction, HSIC was used to design a supervised dictionary learning (SDL) method by maximizing the dependency between the scans taken from “pre-” and “mid-treatment” with “dummy labels” given to the scans. For the dissimilarity measure, an HSIC-based metric was employed to effectively measure the changes in parametric maps as an indication of treatment effectiveness. The HSIC-based feature extraction and dissimilarity measure used a kernel function to nonlinearly transform input vectors into a higher dimensional feature space and computed the population means in the new space, where enhanced group separability was ideally obtained. The results of the classification using the developed CAT system indicated an improvement of performance compared to a CAT system with basic features using histogram of intensity.

Deluge in the size and heterogeneity of medical image databases necessitates the need for content based retrieval systems for their efficient organization. In this paper, we propose such a system to retrieve prostate MR images which share similarities in appearance and content with a query image. We introduce deeply learnt hashing forests (DL-HF) for this image retrieval task. DL-HF effectively leverages the semantic descriptiveness of deep learnt Convolutional Neural Networks. This is used in conjunction with hashing forests which are unsupervised random forests. DL-HF hierarchically parses the deep-learnt feature space to encode subspaces with compact binary code words. We propose a similarity preserving feature descriptor called Parts Histogram which is derived from DL-HF. Correlation defined on this descriptor is used as a similarity metric for retrieval from the database. Validations on publicly available multi-center prostate MR image database established the validity of the proposed approach. The proposed method is fully-automated without any user-interaction and is not dependent on any external image standardization like image normalization and registration. This image retrieval method is generalizable and is well-suited for retrieval in heterogeneous databases other imaging modalities and anatomies.

Intensive international efforts are underway towards phenotyping the mouse genome, by knocking out each of its ≈25,000 genes one-by-one for comparative study. With vast amounts of data to analyze, the traditional method using time-consuming histological examination is clearly impractical, leading to an overwhelming demand for some high-throughput phenotyping framework, especially with the employment of biomedical image informatics to efficiently identify phenotypes concerning morphological abnormality. Existing work has either excessively relied on volumetric analytics which is insensitive to phenotypes associated with no severe volume variations, or tailored for specific defects and thus fails to serve a general phenotyping purpose. Furthermore, the prevailing requirement of an atlas for image segmentation in contrast to its limited availability further complicates the issue in practice. In this paper we propose a high-throughput general-purpose phenotyping framework that is able to efficiently perform batch-wise anomaly detection without prior knowledge of the phenotype and the need for atlas-based segmentation. Anomaly detection is centered on the combined use of group-wise non-rigid image registration and robust principal component analysis (RPCA) for feature extraction and decomposition.

T1-weighted magnetic resonance imaging (MRI) generates contrasts with primary sensitivity to local T1 properties (with lesser T2 and PD contributions). The observed signal intensity is determined by these local properties and the sequence parameters of the acquisition. In common practice, a range of acceptable parameters is used to ensure “similar” contrast across scanners used for any particular study (e.g., the ADNI standard MPRAGE). However, different studies may use different ranges of parameters and report the derived data as simply “T1-weighted”. Physics and imaging authors pay strong heed to the specifics of the imaging sequences, but image processing authors have historically been more lax. Herein, we consider three T1-weighted sequences acquired the same underlying protocol (MPRAGE) and vendor (Philips), but “normal study-to-study variation” in parameters. We show that the gray matter/white matter/cerebrospinal fluid contrast is subtly but systemically different between these images and yields systemically different measurements of brain volume. The problem derives from the visually apparent boundary shifts, which would also be seen by a human rater. We present and evaluate two solutions to produce consistent segmentation results across imaging protocols. First, we propose to acquire multiple sequences on a subset of the data and use the multi-modal imaging as atlases to segment target images any of the available sequences. Second (if additional imaging is not available), we propose to synthesize atlases of the target imaging sequence and use the synthesized atlases in place of atlas imaging data. Both approaches significantly improve consistency of target labeling.

In radiotherapy treatment planning that is only based on magnetic resonance imaging (MRI), the electron density information usually obtained from computed tomography (CT) must be derived from the MRI by synthesizing a so-called pseudo CT (pCT). This is a non-trivial task since MRI intensities are neither uniquely nor quantitatively related to electron density. Typical approaches involve either a classification or regression model requiring specialized MRI sequences to solve intensity ambiguities, or an atlas-based model necessitating multiple registrations between atlases and subject scans. In this work, we explore a machine learning approach for creating a pCT of the pelvic region from conventional MRI sequences without using atlases. We use a random forest provided with information about local texture, edges and spatial features derived from the MRI. This helps to solve intensity ambiguities. Furthermore, we use the concept of auto-context by sequentially training a number of classification forests to create and improve context features, which are finally used to train a regression forest for pCT prediction. We evaluate the pCT quality in terms of the voxel-wise error and the radiologic accuracy as measured by water-equivalent path lengths. We compare the performance of our method against two baseline pCT strategies, which either set all MRI voxels in the subject equal to the CT value of water, or in addition transfer the bone volume from the real CT. We show an improved performance compared to both baseline pCTs suggesting that our method may be useful for MRI-only radiotherapy.

In this work we present a method for dense reconstruction of anatomical structures using white light endoscopic imagery based on a learning process that estimates a mapping between light reflectance and surface geometry. Our method is unique in that few unrealistic assumptions are considered (i.e., we do not assume a Lambertian reflectance model nor do we assume a point light source) and we learn a model on a per-patient basis, thus increasing the accuracy and extensibility to different endoscopic sequences. The proposed method assumes accurate video-CT registration through a combination of Structure-from-Motion (SfM) and Trimmed-ICP, and then uses the registered 3D structure and motion to generate training data with which to learn a multivariate regression of observed pixel values to known 3D surface geometry. We demonstrate with a non-linear regression technique using a neural network towards estimating depth images and surface normal maps, resulting in high-resolution spatial 3D reconstructions to an average error of 0.53mm (on the low side, when anatomy matches the CT precisely) to 1.12mm (on the high side, when the presence of liquids causes scene geometry that is not present in the CT for evaluation). Our results are exhibited on patient data and validated with associated CT scans. In total, we processed 206 total endoscopic images from patient data, where each image yields approximately 1 million reconstructed 3D points per image.

We introduce a framework for quantitative evaluation of 3D vessel segmentation approaches using vascular phantoms. Phantoms are designed using a CAD system and created with a 3D printer, and comprise realistic shapes including branches and pathologies such as abdominal aortic aneurysms (AAA). To transfer ground truth information to the 3D image coordinate system, we use a landmark-based registration scheme utilizing fiducial markers integrated in the phantom design. For accurate 3D localization of the markers we developed a novel 3D parametric intensity model that is directly fitted to the markers in the images. We also performed a quantitative evaluation of different vessel segmentation approaches for a phantom of an AAA.

In a fetal brain screening examination, a standardized set of anatomical views is inspected and certain biometric measurements are taken in these views. Acquisition of recommended planes requires a certain level of operator expertise. 3D ultrasound has the potential to reduce the manual task to only capture a volume containing the head and to subsequently determine the standard 2D views and measurements automatically. For this purpose, a segmentation model of the fetal brain was created and trained with expert annotations. It was found that the annotations show a considerable intra- and inter-observer variability. To handle the variability, we propose a method to train the model with redundant but inconsistent reference data from many expert users. If the outlier-cleaned average of all reference annotations is considered as ground truth, errors of the automatic view detection are lower than the errors of all individual users and errors of the measurements are in the same range as user error. The resulting functionality allows the completely automated estimation of views and measurements in 3D fetal ultrasound images.

Segmentation of breast tissue in MRI images is an important pre-processing step for many applications. We present a new method that uses a random forest classifier to identify candidate edges in the image and then applies a Poisson reconstruction step to define a 3D surface based on the detected edge points. Using a leave one patient out cross validation we achieve a Dice overlap score of 0.96 ± 0.02 for T1 weighted non-fat suppressed images in 8 patients. In a second dataset of 332 images acquired using a Dixon sequence, which was not used in training the random classifier, the mean Dice score was 0.90 ± 0.03. Using this approach we have achieved accurate, robust segmentation results using a very small training set.

Optical coherence tomography (OCT) is a noninvasive imaging modality that has begun to find widespread use in retinal imaging for the detection of a variety of ocular diseases. In addition to structural changes in the form of altered retinal layer thicknesses, pathological conditions may also cause the formation of edema within the retina. In multiple sclerosis, for instance, the nerve fiber and ganglion cell layers are known to thin. Additionally, the formation of pseudocysts called microcystic macular edema (MME) have also been observed in the eyes of about 5% of MS patients, and its presence has been shown to be correlated with disease severity. Previously, we proposed separate algorithms for the segmentation of retinal layers and MME, but since MME mainly occurs within specific regions of the retina, a simultaneous approach is advantageous. In this work, we propose an automated globally optimal graph-theoretic approach that simultaneously segments the retinal layers and the MME in volumetric OCT scans. SD-OCT scans from one eye of 12 MS patients with known MME and 8 healthy controls were acquired and the pseudocysts manually traced. The overall precision and recall of the pseudocyst detection was found to be 86.0% and 79.5%, respectively.

We present a fully automated generative method for simultaneous brain tumor and organs-at-risk segmentation in multi-modal magnetic resonance images. The method combines an existing whole-brain segmentation technique with a spatial tumor prior, which uses convolutional restricted Boltzmann machines to model tumor shape. The method is not tuned to any specific imaging protocol and can simultaneously segment the gross tumor volume, peritumoral edema and healthy tissue structures relevant for radiotherapy planning. We validate the method on a manually delineated clinical data set of glioblastoma patients by comparing segmentations of gross tumor volume, brainstem and hippocampus. The preliminary results demonstrate the feasibility of the method.

Accurate and reproducible segmentation of cysts and fluid-filled regions from retinal OCT images is an important step allowing quantification of the disease status, longitudinal disease progression, and response to therapy in wet-pathology retinal diseases. However, segmentation of fluid-filled regions from OCT images is a challenging task due to their inhomogeneous appearance, the unpredictability of their number, size and location, as well as the intensity profile similarity between such regions and certain healthy tissue types. While machine learning techniques can be beneficial for this task, they require large training datasets and are often over-fitted to the appearance models of specific scanner vendors. We propose a knowledge-based approach that leverages a carefully designed cost function and graph-based segmentation techniques to provide a vendor-independent solution to this problem. We illustrate the results of this approach on two publicly available datasets with a variety of scanner vendors and retinal disease status. Compared to a previous machine-learning based approach, the volume similarity error was dramatically reduced from 81:3±56:4% to 22:2±21:3% (paired t-test, p << 0:001).

The optic disc (OD) and the macula are important structures in automatic diagnosis of most retinal diseases inducing vision defects such as glaucoma, diabetic or hypertensive retinopathy and age-related macular degeneration. We propose a new method to detect simultaneously the macula and the OD boundary. First, the color fundus images are processed to compute several maps highlighting the different anatomical structures such as vessels, the macula and the OD. Then, macula candidates and OD candidates are found simultaneously and independently using seed detectors identified on the corresponding maps. After selecting a set of macula/OD pairs, the top candidates are sent to the OD segmentation method. The segmentation method is based on local K-means applied to color coordinates in polar space followed by a polynomial fitting regularization step. Pair scores are updated, resulting in the final best macula/OD pair. The method was evaluated on two public image databases: ONHSD and MESSIDOR. The results show an overlapping area of 0.84 on ONHSD and 0.90 on MESSIDOR, which is better than recent state of the art methods. Our segmentation method is robust to contrast and illumination problems and outputs the exact boundary of the OD, not just a circular or elliptical model. The macula detection has an accuracy of 94%, which again outperforms other macula detection methods. This shows that combining the OD and macula detections improves the overall accuracy. The computation time for the whole process is 6.4 seconds, which is faster than other methods in the literature.

Pathologies of the optic nerve and orbit impact millions of Americans and quantitative assessment of the orbital structures on 3-D imaging would provide objective markers to enhance diagnostic accuracy, improve timely intervention, and eventually preserve visual function. Recent studies have shown that the multi-atlas methodology is suitable for identifying orbital structures, but challenges arise in the identification of the individual extraocular rectus muscles that control eye movement. This is increasingly problematic in diseased eyes, where these muscles often appear to fuse at the back of the orbit (at the resolution of clinical computed tomography imaging) due to inflammation or crowding. We propose the use of Kalman filters to track the muscles in three-dimensions to refine multi-atlas segmentation and resolve ambiguity due to imaging resolution, noise, and artifacts. The purpose of our study is to investigate a method of automatically generating orbital metrics from CT imaging and demonstrate the utility of the approach by correlating structural metrics of the eye orbit with clinical data and visual function measures in subjects with thyroid eye disease. The pilot study demonstrates that automatically calculated orbital metrics are strongly correlated with several clinical characteristics. Moreover, it is shown that the superior, inferior, medial and lateral rectus muscles obtained using Kalman filters are each correlated with different categories of functional deficit. These findings serve as foundation for further investigation in the use of CT imaging in the study, analysis and diagnosis of ocular diseases, specifically thyroid eye disease.

Cerebellar peduncles (CPs) are white matter tracts connecting the cerebellum to other brain regions. Automatic segmentation methods of the CPs have been proposed for studying their structure and function. Usually the performance of these methods is evaluated by comparing segmentation results with manual delineations (ground truth). However, when a segmentation method is run on new data (for which no ground truth exists) it is highly desirable to efficiently detect and assess algorithm failures so that these cases can be excluded from scientific analysis. In this work, two outlier detection methods aimed to assess the performance of an automatic CP segmentation algorithm are presented. The first one is a univariate non-parametric method using a box-whisker plot. We first categorize automatic segmentation results of a dataset of diffusion tensor imaging (DTI) scans from 48 subjects as either a success or a failure. We then design three groups of features from the image data of nine categorized failures for failure detection. Results show that most of these features can efficiently detect the true failures. The second method—supervised classification—was employed on a larger DTI dataset of 249 manually categorized subjects. Four classifiers—linear discriminant analysis (LDA), logistic regression (LR), support vector machine (SVM), and random forest classification (RFC)—were trained using the designed features and evaluated using a leave-one-out cross validation. Results show that the LR performs worst among the four classifiers and the other three perform comparably, which demonstrates the feasibility of automatically detecting segmentation failures using classification methods.

Current fully automatic liver tumor segmentation systems are designed to work on a single CT-image. This hinders these systems from the detection of more complex types of liver tumor. We therefore present a new algorithm for liver tumor segmentation that allows incorporating different CT scans and requires no manual interaction. We derive a liver segmentation with state-of-the-art shape models which are robust to initialization. The tumor segmentation is then achieved by classifying all voxels into healthy or tumorous tissue using Extremely Randomized Trees with an auto-context learning scheme. Using DALSA enables us to learn from only sparse annotations and allows a fast set-up for new image settings. We validate the quality of our algorithm with exemplary segmentation results.

This paper addresses the localization of anatomical structures in medical images by a Generalized Hough Transform (GHT). As localization is often a pre-requisite for subsequent model-based segmentation, it is important to assess whether or not the GHT was able to locate the desired object. The GHT by its construction does not make this distinction. We present an approach to detect incorrect GHT localizations by deriving collective features of contributing GHT model points and by training a Support Vector Machine (SVM) classifier. On a training set of 204 cases, we demonstrate that for the detection of incorrect localizations classification errors of down to 3% are achievable. This is three times less than the observed intrinsic GHT localization error.

Sparse learning enables dimension reduction and efficient modeling of high dimensional signals and images, but it may need to be tailored to best suit specific applications and datasets. Here we used sparse learning to efficiently represent functional magnetic resonance imaging (fMRI) data from the human brain. We propose a novel embedded sparse representation (ESR), to identify the most consistent dictionary atoms across different brain datasets via an iterative group-wise dictionary optimization procedure. In this framework, we introduced additional criteria to make the learned dictionary atoms more consistent across different subjects. We successfully identified four common dictionary atoms that follow the external task stimuli with very high accuracy. After projecting the corresponding coefficient vectors back into the 3-D brain volume space, the spatial patterns are also consistent with traditional fMRI analysis results. Our framework reveals common features of brain activation in a population, as a new, efficient fMRI analysis method.

Arterial calcification has been related to cardiovascular disease (CVD) and osteoporosis. However, little is known about the role of genetics and exact pathways leading to arterial calcification and its relation to bone density changes indicating osteoporosis. In this study, we conducted a genome-wide association study of arterial calcification burden, followed by a look-up of known single nucleotide polymorphisms (SNPs) for coronary artery disease (CAD) and myocardial infarction (MI), and bone mineral density (BMD) to test for a shared genetic basis between the traits. The study included a subcohort of the Dutch-Belgian lung cancer screening trial comprised of 2,561 participants. Participants underwent baseline CT screening in one of two hospitals participating in the trial. Low-dose chest CT images were acquired without contrast enhancement and without ECG-synchronization. In these images coronary and aortic calcifications were identified automatically. Subsequently, the detected calcifications were quantified using coronary artery calcium Agatston and volume scores. Genotype data was available for these participants. A genome-wide association study was conducted on 10,220,814 SNPs using a linear regression model. To reduce multiple testing burden, known CAD/MI and BMD SNPs were specifically tested (45 SNPs from the CARDIoGRAMplusC4D consortium and 60 SNPS from the GEFOS consortium). No novel significant SNPs were found. Significant enrichment for CAD/MI SNPs was observed in testing Agatston and coronary artery calcium volume scores. Moreover, a significant enrichment of BMD SNPs was shown in aortic calcium volume scores. This may indicate genetic relation of BMD SNPs and arterial calcification burden.

Cochlear Implants (CIs) are electrode arrays that are surgically inserted into the cochlea. Individual contacts stimulate frequency-mapped nerve endings thus replacing the natural electro-mechanical transduction mechanism. CIs are programmed post-operatively by audiologists but this is currently done using behavioral tests without imaging information that permits relating electrode position to inner ear anatomy. We have recently developed a series of image processing steps that permit the segmentation of the inner ear anatomy and the localization of individual contacts. We have proposed a new programming strategy that uses this information and we have shown in a study with 68 participants that 78% of long term recipients preferred the programming parameters determined with this new strategy. A limiting factor to the large scale evaluation and deployment of our technique is the amount of user interaction still required in some of the steps used in our sequence of image processing algorithms. One such step is the rough registration of an atlas to target volumes prior to the use of automated intensity-based algorithms when the target volumes have very different fields of view and orientations. In this paper we propose a solution to this problem. It relies on a random forest-based approach to automatically localize a series of landmarks. Our results obtained from 83 images with 132 registration tasks show that automatic initialization of an intensity-based algorithm proves to be a reliable technique to replace the manual step.

Percutaneous radio frequency ablation is a method for liver tumor treatment when conventional surgery is not an option. It is a minimally invasive treatment and may be performed under CT image guidance if the tumor does not give sufficient contrast on ultrasound images. For optimal guidance, registration of the pre-operative contrast-enhanced CT image to the intra-operative CT image is hypothesized to improve guidance. This is a highly challenging registration task due to large differences in pose and image quality. In this study, we introduce a semi-automated registration algorithm to address this problem. The method is based on a conventional nonrigid intensity-based registration framework, extended with a novel point-to-surface constraint. The point-to-surface constraint serves to improve the alignment of the liver boundary, while requiring minimal user interaction during the operation. The method assumes that a liver segmentation of the pre-operative CT is available. After an initial nonrigid registration without the point-to-surface constraint, the operator clicks a few points on the liver surface at those regions where the nonrigid registration seems inaccurate. In a subsequent registration step, these points on the intra-operative image are driven towards the liver surface on the preoperative image, using a penalty term added to the registration cost function. The method is evaluated on five clinical datasets and it is shown to improve registration compared with conventional rigid and nonrigid registrations in all cases.

Identifying cross-sectional and longitudinal correspondence in the abdomen on computed tomography (CT) scans is necessary for quantitatively tracking change and understanding population characteristics, yet abdominal image registration is a challenging problem. The key difficulty in solving this problem is huge variations in organ dimensions and shapes across subjects. The current standard registration method uses the global or body-wise registration technique, which is based on the global topology for alignment. This method (although producing decent results) has substantial influence of outliers, thus leaving room for significant improvement. Here, we study a new image registration approach using local (organ-wise registration) by first creating organ-specific bounding boxes and then using these regions of interest (ROIs) for aligning references to target. Based on Dice Similarity Coefficient (DSC), Mean Surface Distance (MSD) and Hausdorff Distance (HD), the organ-wise approach is demonstrated to have significantly better results by minimizing the distorting effects of organ variations. This paper compares exclusively the two registration methods by providing novel quantitative and qualitative comparison data and is a subset of the more comprehensive problem of improving the multi-atlas segmentation by using organ normalization.

We introduce an efficient and novel convex optimization-based approach to the challenging non-rigid registration of 3D prostate magnetic resonance (MR) and transrectal ultrasound (TRUS) images, which incorporates a new volume preserving constraint to essentially improve the accuracy of targeting suspicious regions during the 3D TRUS guided prostate biopsy. Especially, we propose a fast sequential convex optimization scheme to efficiently minimize the employed highly nonlinear image fidelity function using the robust multi-channel modality independent neighborhood descriptor (MIND) across the two modalities of MR and TRUS. The registration accuracy was evaluated using 10 patient images by calculating the target registration error (TRE) using manually identified corresponding intrinsic fiducials in the whole prostate gland. We also compared the MR and TRUS manually segmented prostate surfaces in the registered images in terms of the Dice similarity coefficient (DSC), mean absolute surface distance (MAD), and maximum absolute surface distance (MAXD). Experimental results showed that the proposed method with the introduced volume-preserving prior significantly improves the registration accuracy comparing to the method without the volume-preserving constraint, by yielding an overall mean TRE of 2:0±0:7 mm, and an average DSC of 86:5±3:5%, MAD of 1:4±0:6 mm and MAXD of 6:5±3:5 mm.

The number of minimally invasive procedures is growing every year. These procedures are highly complex and very demanding for the surgeons. It is therefore important to provide intraoperative assistance to alleviate these difficulties. For most computer-assistance systems, like visualizing target structures with augmented reality, a registration step is required to map preoperative data (e.g. CT images) to the ongoing intraoperative scene. Without additional hardware, the (stereo-) endoscope is the prime intraoperative data source and with it, stereo reconstruction methods can be used to obtain 3D models from target structures. To link reconstructed parts from different frames (mosaicking), the endoscope movement has to be known. In this paper, we present a camera tracking method that uses dense depth and feature registration which are combined with a Kalman Filter scheme. It provides a robust position estimation that shows promising results in ex vivo and in silico experiments.

Nonlinear registration algorithms are routinely used in brain imaging, to align data for inter-subject and group comparisons, and for voxelwise statistical analyses. To understand how the choice of registration method affects maps of functional brain connectivity in a sample of 611 twins, we evaluated three popular nonlinear registration methods: Advanced Normalization Tools (ANTs), Automatic Registration Toolbox (ART), and FMRIB's Nonlinear Image Registration Tool (FNIRT). Using both structural and functional MRI, we used each of the three methods to align the MNI152 brain template, and 80 regions of interest (ROIs), to each subject's T1-weighted (T1w) anatomical image. We then transformed each subject's ROIs onto the associated resting state functional MRI (rs-fMRI) scans and computed a connectivity network or functional connectome for each subject. Given the different degrees of genetic similarity between pairs of monozygotic (MZ) and same-sex dizygotic (DZ) twins, we used structural equation modeling to estimate the additive genetic influences on the elements of the function networks, or their heritability. The functional connectome and derived statistics were relatively robust to nonlinear registration effects.

Classification of clustered breast microcalcifications into benign and malignant categories is an extremely challenging task for computerized algorithms and expert radiologists alike. In this paper we present a novel method for feature selection based on mutual information (MI) criterion for automatic classification of microcalcifications. We explored the MI based feature selection for various texture features. The proposed method was evaluated on a standardized digital database for screening mammography (DDSM). Experimental results demonstrate the effectiveness and the advantage of using the MI-based feature selection to obtain the most relevant features for the task and thus to provide for improved performance as compared to using all features.

Recently, several studies have successfully applied multivariate pattern analysis methods to predict the categories of emotions. These studies are mainly focused on self-experienced emotions, such as the emotional states elicited by music or movie. In fact, most of our social interactions involve perception of emotional information from the expressions of other people, and it is an important basic skill for humans to recognize the emotional facial expressions of other people in a short time. In this study, we aimed to determine the discriminability of perceived emotional facial expressions. In a rapid event-related fMRI design, subjects were instructed to classify four categories of facial expressions (happy, disgust, angry and neutral) by pressing different buttons, and each facial expression stimulus lasted for 2s. All participants performed 5 fMRI runs. One multivariate pattern analysis method, support vector machine was trained to predict the categories of facial expressions. For feature selection, ninety masks defined from anatomical automatic labeling (AAL) atlas were firstly generated and each were treated as the input of the classifier; then, the most stable AAL areas were selected according to prediction accuracies, and comprised the final feature sets. Results showed that: for the 6 pair-wise classification conditions, the accuracy, sensitivity and specificity were all above chance prediction, among which, happy vs. neutral , angry vs. disgust achieved the lowest results. These results suggested that specific neural signatures of perceived emotional facial expressions may exist, and happy vs. neutral, angry vs. disgust might be more similar in information representation in the brain.

Clean bone segmentation is critical in studying the joint anatomy for measuring the spacing between the bones. However, separation of the coupled bones in CT images is sometimes difficult due to ambiguous gray values coming from the noise and the heterogeneity of bone materials as well as narrowing of the joint space. For fine reconstruction of the individual local boundaries, manual operation is a common practice where the segmentation remains to be a bottleneck. In this paper, we present an automatic method for extracting the joint space by applying graph cut on Markov random field model to the region of interest (ROI) which is identified by a template of 3D bone structures. The template includes encoded articular surface which identifies the tight region of the high-intensity bone boundaries together with the fuzzy joint area of interest. The localized shape information from the template model within the ROI effectively separates the bones nearby. By narrowing the ROI down to the region including two types of tissue, the object extraction problem was reduced to binary segmentation and solved via graph cut. Based on the shape of a joint space marked by the template, the hard constraint was set by the initial seeds which were automatically generated from thresholding and morphological operations. The performance and the robustness of the proposed method are evaluated on 12 volumes of ankle CT data, where each volume includes a set of 4 tarsal bones (calcaneus, talus, navicular and cuboid).

This paper describes enhancements to automate classification of brain tissues for multi-site degenerative magnetic resonance imaging (MRI) data analysis. Processing of large collections of MR images is a key research technique to advance our understanding of the human brain. Previous studies have developed a robust multi-modal tool for automated tissue classification of large-scale data based on expectation maximization (EM) method initialized by group-wise prior probability distributions. This work aims to augment the EM-based classification using a non-parametric fuzzy k-Nearest Neighbor (k-NN) classifier that can model the unique anatomical states of each subject in the study of degenerative diseases. The presented method is applicable to multi-center heterogeneous data analysis and is quantitatively validated on a set of 18 synthetic multi-modal MR datasets having six different levels of noise and three degrees of bias-field provided with known ground truth. Dice index and average Hausdorff distance are used to compare the accuracy and robustness of the proposed method to a state-of-the-art classification method implemented based on EM algorithm. Both evaluation measurements show that presented enhancements produce superior results as compared to the EM only classification.

Pancreatic ductal adenocarcinoma (PDAC) is the fourth leading cause of cancer-related death in the United States. The five-year survival rate for all stages is approximately 6%, and approximately 2% when presenting with distant disease.¹ Only 10-20% of all patients present with resectable disease, but recurrence rates are high with only 5 to 15% remaining free of disease at 5 years. At this time, we are unable to distinguish between resectable PDAC patients with occult metastatic disease from those with potentially curable disease. Early classification of these tumor types may eventually lead to changes in initial management including the use of neoadjuvant chemotherapy or radiation, or in the choice of postoperative adjuvant treatments. Texture analysis is an emerging methodology in oncologic imaging for quantitatively assessing tumor heterogeneity that could potentially aid in the stratification of these patients. The present study derives several texture-based features from CT images of PDAC patients, acquired prior to neoadjuvant chemotherapy, and analyzes their performance, individually as well as in combination, as prognostic markers. A fuzzy minimum redundancy maximum relevance method with leave-one-image-out technique is included to select discriminating features from the set of extracted features. With a naive Bayes classifier, the proposed method predicts the 5-year overall survival of PDAC patients prior to neoadjuvant therapy and achieves the best results in terms of the area under the receiver operating characteristic curve of 0:858 and accuracy of 83:0% with four-fold cross-validation techniques.

Osteoporosis is the common cause for a broken bone among senior citizens. Early diagnosis of osteoporosis requires routine examination which may be costly for patients. A potential low cost diagnosis is to identify a senior citizen at high risk of osteoporosis by pre-screening during routine dental examination. Therefore, osteoporosis analysis using dental radiographs severs as a key step in routine dental examination. The aim of this study is to localize landmarks in dental radiographs which are helpful to assess the evidence of osteoporosis. We collect eight landmarks which are critical in osteoporosis analysis. Our goal is to localize these landmarks automatically for a given dental radiographic image. To address the challenges such as large variations of appearances in subjects, in this paper, we formulate the task into a multi-class classification problem. A hybrid feature pool is used to represent these landmarks. For the discriminative classification problem, we use a random forest to fuse the hybrid feature representation. In the experiments, we also evaluate the performances of individual feature component and the hybrid fused feature. Our proposed method achieves average detection error of 2:9mm.

Localization of anatomical regions of interest (ROIs) is a preprocessing step in many medical image analysis tasks. While trivial for humans, it is complex for automatic methods. Classic machine learning approaches require the challenge of hand crafting features to describe differences between ROIs and background. Deep convolutional neural networks (CNNs) alleviate this by automatically finding hierarchical feature representations from raw images. We employ this trait to detect anatomical ROIs in 2D image slices in order to localize them in 3D.

In 100 low-dose non-contrast enhanced non-ECG synchronized screening chest CT scans, a reference standard was defined by manually delineating rectangular bounding boxes around three anatomical ROIs — heart, aortic arch, and descending aorta. Every anatomical ROI was automatically identified using a combination of three CNNs, each analyzing one orthogonal image plane. While single CNNs predicted presence or absence of a specific ROI in the given plane, the combination of their results provided a 3D bounding box around it.

Classification performance of each CNN, expressed in area under the receiver operating characteristic curve, was ≥0.988. Additionally, the performance of ROI localization was evaluated. Median Dice scores for automatically determined bounding boxes around the heart, aortic arch, and descending aorta were 0.89, 0.70, and 0.85 respectively. The results demonstrate that accurate automatic 3D localization of anatomical structures by CNN-based 2D image classification is feasible.

Body weight adapted drug dosages are important for emergency treatments: Inaccuracies in body weight estimation may lead to inaccurate drug dosing. This paper describes an improved approach to estimating the body weight of emergency patients in a trauma room, based on images from an RGB-D and a thermal camera. The improvements are specific to several aspects: Fusion of RGB-D and thermal camera eases filtering and segmentation of the patient's body from the background. Robustness and accuracy is gained by an artificial neural network, which considers geometric features from the sensors as input, e.g. the patient's volume, and shape parameters. Preliminary experiments with 69 patients show an accuracy close to 90 percent, with less than 10 percent relative error and the results are compared with the physician's estimate, the patient's statement and an established anthropometric method.

Computed tomography is the modality of choice for poly-trauma patients to assess rapidly skeletal and vascular integrity of the whole body. Often several scans with and without contrast medium or with different spatial resolution are acquired. Efficient reading of the resulting extensive set of image data is vital, since it is often time critical to initiate the necessary therapeutic actions. A set of automatically found landmarks can facilitate navigation in the data and enables anatomy oriented viewing. Following this intention, we selected a comprehensive set of 17 skeletal and 5 aortic landmarks. Landmark localization models for the Discriminative Generalized Hough Transform (DGHT) were automatically created based on a set of about 20 training images with ground truth landmark positions. A hierarchical setup with 4 resolution levels was used. Localization results were evaluated on a separate test set, consisting of 50 to 128 images (depending on the landmark) with available ground truth landmark locations. The image data covers a large amount of variability caused by differences of field-of-view, resolution, contrast agent, patient gender and pathologies. The median localization error for the set of aortic landmarks was 14.4 mm and for the set of skeleton landmarks 5.5 mm. Median localization errors for individual landmarks ranged from 3.0 mm to 31.0 mm. The runtime performance for the whole landmark set is about 5s on a typical PC.

Novelty detection is concerned with identifying test data that differs from the training data of a classifier. In the case of brain MR images, pathology or imaging artefacts are examples of untrained data. In this proof-of-principle study, we measure the behaviour of a classifier during the classification of trained labels (i.e. normal brain tissue). Next, we devise a measure that distinguishes normal classifier behaviour from abnormal behavior that occurs in the case of a novelty. This will be evaluated by training a kNN classifier on normal brain tissue, applying it to images with an untrained pathology (white matter hyperintensities (WMH)), and determine if our measure is able to identify abnormal classifier behaviour at WMH locations. For our kNN classifier, behaviour is modelled as the mean, median, or q1 distance to the k nearest points. Healthy tissue was trained on 15 images; classifier behaviour was trained/tested on 5 images with leave-one-out cross-validation. For each trained class, we measure the distribution of mean/median/q1 distances to the k nearest point. Next, for each test voxel, we compute its Z-score with respect to the measured distribution of its predicted label. We consider a Z-score ≥4 abnormal behaviour of the classifier, having a probability due to chance of 0.000032. Our measure identified >90% of WMH volume and also highlighted other non-trained findings. The latter being predominantly vessels, cerebral falx, brain mask errors, choroid plexus. This measure is generalizable to other classifiers and might help in detecting unexpected findings or novelties by measuring classifier behaviour.

Early detection of risk is critical in determining the course of treatment in traumatic brain injury (TBI). Computed tomography (CT) acquired at admission has shown latent prognostic value in prior studies; however, no robust clinical risk predictions have been achieved based on the imaging data in large-scale TBI analysis. The major challenge lies in the lack of consistent and complete medical records for patients, and an inherent bias associated with the limited number of patients samples with high-risk outcomes in available TBI datasets. Herein, we propose a Bayesian framework with mutual information-based forward feature selection to handle this type of data. Using multi-atlas segmentation, 154 image-based features (capturing intensity, volume and texture) were computed over 22 ROIs in 1791 CT scans. These features were combined with 14 clinical parameters and converted into risk likelihood scores using Bayes modeling. We explore the prediction power of the image features versus the clinical measures for various risk outcomes. The imaging data alone were more predictive of outcomes than the clinical data (including Marshall CT classification) for discharge disposition with an area under the curve of 0.81 vs. 0.67, but less predictive than clinical data for discharge Glasgow Coma Scale (GCS) score with an area under the curve of 0.65 vs. 0.85. However, in both cases, combining imaging and clinical data increased the combined area under the curve with 0.86 for discharge disposition and 0.88 for discharge GCS score. In conclusion, CT data have meaningful prognostic value for TBI patients beyond what is captured in clinical measures and the Marshall CT classification.

It has been recently proved that the computational analysis of X-ray Computed Tomography (CT) images allows clinicians to assess the alteration of compact bone asset due to hematological diseases. HT-BONE implements a new method, based on an extension of the Hough transform (HT) to a wide class of algebraic curves, for accurately measuring global and regional geometric properties of trabecular and compact bone districts. In the case of CT/PET analysis, the segmentation of the CT images provides masks for Positron Emission Tomography (PET) data, extracting the metabolic activity in the region surrounded by compact bone tissue. HT-BONE offers an intuitive, user-friendly, Matlab-based Graphical User Interface (GUI) for all input/output procedures and the automatic managing of the segmentation process also from non-expert users: the CT/PET data can be loaded and browsed easily and the only pre-preprocessing required from the user is the drawing of Regions Of Interest (ROIs) around the bone districts under consideration. For each bone district, specific families of curves, whose reliability has been already tested in previous works, is automatically selected for the recognition task via HT. As output, the software returns masks of the segmented compact bone regions, images of the Standard Uptake Values (SUV) in the masked regions of PET slices, and the values of the parameters in the curve equations utilized in the HT procedure. This information can be used for all pathologies and clinical conditions for which the alteration of the compact bone asset or bone marrow distribution plays a crucial role.

Glioblastoma multiforme (GBM) is the most common malignant primary tumor of the central nervous system, characterized among other traits by rapid metastatis. Three tissue phenotypes closely associated with GBMs, namely, necrosis (N), contrast enhancement (CE), and edema/invasion (E), exhibit characteristic patterns of texture heterogeneity in magnetic resonance images (MRI). In this study, we propose a novel model to characterize GBM tissue phenotypes using gray level co-occurrence matrices (GLCM) in three anatomical planes. The GLCM encodes local image patches in terms of informative, orientation-invariant texture descriptors, which are used here to sub-classify GBM tissue phenotypes. Experiments demonstrate the model on MRI data of 41 GBM patients, obtained from the cancer genome atlas (TCGA). Intensity-based automatic image registration is applied to align corresponding pairs of fixed T1˗weighted (T1˗WI) post-contrast and fluid attenuated inversion recovery (FLAIR) images. GBM tissue regions are then segmented using the 3D Slicer tool. Texture features are computed from 12 quantifier functions operating on GLCM descriptors, that are generated from MRI intensities within segmented GBM tissue regions. Various classifier models are used to evaluate the effectiveness of texture features for discriminating between GBM phenotypes. Results based on T1-WI scans showed a phenotype classification accuracy of over 88.14%, a sensitivity of 85.37% and a specificity of 96.1%, using the linear discriminant analysis (LDA) classifier. This model has the potential to provide important characteristics of tumors, which can be used for the sub-classification of GBM phenotypes.

Branch retinal artery occlusion (BRAO) is an ocular emergency which could lead to blindness. Quantitative analysis of BRAO region in the retina is very needed to assessment of the severity of retinal ischemia. In this paper, a fully automatic framework was proposed to classify and segment BRAO based on 3D spectral-domain optical coherence tomography (SD-OCT) images. To the best of our knowledge, this is the first automatic 3D BRAO segmentation framework. First, a support vector machine (SVM) based classifier is designed to differentiate BRAO into acute phase and chronic phase, and the two types are segmented separately. To segment BRAO in chronic phase, a threshold-based method is proposed based on the thickness of inner retina. While for segmenting BRAO in acute phase, a two-step segmentation is performed, which includes the bayesian posterior probability based initialization and the graph-search-graph-cut based segmentation. The proposed method was tested on SD-OCT images of 23 patients (12 of acute and 11 of chronic phase) using leave-one-out strategy. The overall classification accuracy of SVM classifier was 87.0%, and the TPVF and FPVF for acute phase were 91.1%, 5.5%; for chronic phase were 90.5%, 8.7%, respectively.

In this work some segmentation techniques are evaluated by using a simple centroid-based classification system regarding breast mass delineation in digital mammography images. The aim is to determine the best one for future CADx developments. Six techniques were tested: Otsu, SOM, EICAMM, Fuzzy C-Means, K-Means and Level-Set. All of them were applied to segment 317 mammography images from DDSM database. A single compact set of attributes was extracted and two centroids were defined, one for malignant and another for benign cases. The final classification was based on proximity with a given centroid and the best results were presented by the Level-Set technique with a 68.1% of Accuracy, which indicates this method as the most promising for breast masses segmentation aiming a more precise interpretation in schemes CADx.

Prostate segmentation on CT images is a challenging task. In this paper, we explore the population and patient-specific characteristics for the segmentation of the prostate on CT images. Because population learning does not consider the inter-patient variations and because patient-specific learning may not perform well for different patients, we are combining the population and patient-specific information to improve segmentation performance. Specifically, we train a population model based on the population data and train a patient-specific model based on the manual segmentation on three slice of the new patient. We compute the similarity between the two models to explore the influence of applicable population knowledge on the specific patient. By combining the patient-specific knowledge with the influence, we can capture the population and patient-specific characteristics to calculate the probability of a pixel belonging to the prostate. Finally, we smooth the prostate surface according to the prostate-density value of the pixels in the distance transform image. We conducted the leave-one-out validation experiments on a set of CT volumes from 15 patients. Manual segmentation results from a radiologist serve as the gold standard for the evaluation. Experimental results show that our method achieved an average DSC of 85.1% as compared to the manual segmentation gold standard. This method outperformed the population learning method and the patient-specific learning approach alone. The CT segmentation method can have various applications in prostate cancer diagnosis and therapy.

The development of an accurate, robust and fast segmentation algorithm has long been a research focus in medical computer vision. State-of-the-art practices often involve non-rigidly registering a target image with a set of training atlases for label propagation over the target space to perform segmentation, a.k.a. multi-atlas label propagation (MALP). In recent years, the patch-based segmentation (PBS) framework has gained wide attention due to its advantage of relaxing the strict voxel-to-voxel correspondence to a series of pair-wise patch comparisons for contextual pattern matching. Despite a high accuracy reported in many scenarios, computational efficiency has consistently been a major obstacle for both approaches. Inspired by recent work on random forest, in this paper we propose a patch forest approach, which by equipping the conventional PBS with a fast patch search engine, is able to boost segmentation speed significantly while retaining an equal level of accuracy. In addition, a fast forest training mechanism is also proposed, with the use of a dynamic grid framework to efficiently approximate data compactness computation and a 3D integral image technique for fast box feature retrieval.

Automatic, quick and reliable identification of retinal landmarks from fundus photography is key for measurements used in research, diagnosis, screening and treating of common diseases affecting the eyes. This study presents a fast method for the detection of the centre of mass of the vascular arcades, optic nerve head (ONH) and fovea, used in the definition of five clinically relevant areas in use for screening programmes for diabetic retinopathy (DR). Thirty-eight fundus photographs showing 7203 DR lesions were analysed to find the landmarks manually by two retina-experts and automatically by the proposed method. The automatic identification of the ONH and fovea were performed using template matching based on normalised cross correlation. The centre of mass of the arcades was obtained by fitting an ellipse on sample coordinates of the main vessels. The coordinates were obtained by processing the image with hessian filtering followed by shape analyses and finally sampling the results. The regions obtained manually and automatically were used to count the retinal lesions falling within, and to evaluate the method. 92.7% of the lesions were falling within the same regions based on the landmarks selected by the two experts. 91.7% and 89.0% were counted in the same areas identified by the method and the first and second expert respectively. The inter-repeatability of the proposed method and the experts is comparable, while the 100% intra-repeatability makes the algorithm a valuable tool in tasks like analyses in real-time, of large datasets and of intra-patient variability.

Identification of the basal slice in cardiac imaging is a key step to measuring the ejection fraction (EF) of the left ventricle (LV). Despite research on cardiac segmentation, basal slice identification is routinely performed manually. Manual identification, however, has been shown to have high inter-observer variability, with a variation of the EF by up to 8%. Therefore, an automatic way of identifying the basal slice is still required. Prior published methods operate by automatically tracking the mitral valve points from the long-axis view of the LV. These approaches assumed that the basal slice is the first short-axis slice below the mitral valve. However, guidelines published in 2013 by the society for cardiovascular magnetic resonance indicate that the basal slice is the uppermost short-axis slice with more than 50% myocardium surrounding the blood cavity. Consequently, these existing methods are at times identifying the incorrect short-axis slice. Correct identification of the basal slice under these guidelines is challenging due to the poor image quality and blood movement during image acquisition. This paper proposes an automatic tool that focuses on the two-chamber slice to find the basal slice. To this end, an active shape model is trained to automatically segment the two-chamber view for 51 samples using the leave-one-out strategy. The basal slice was detected using temporal binary profiles created for each short-axis slice from the segmented two-chamber slice. From the 51 successfully tested samples, 92% and 84% of detection results were accurate at the end-systolic and the end-diastolic phases of the cardiac cycle, respectively.

Recent efforts to perform a 3D reconstruction of interventional devices such as guidewires from monoplane and biplane fluoroscopic images require the segmentation of the exact device path in the respective 2D projection images. The segmentation of the device in low dose fluoroscopy images can be challenging since noise and motion artifacts degrade the image quality. Additionally, extracting the device path from the segmented region may result in ambiguous results due to overlapping device parts or discontinuities in the device segmentation. The purpose of this work is to present a novel guidewire tracking and segmentation algorithm, which segments the device region based on three different features based on a ridge detection filter, noise reduction for curvilinear structures as well as an a priori probability map. The features are calculated from background subtracted as well as unsubtracted fluoroscopic images. The device path extraction is based on a topology preserving thinning algorithm followed by a path search, which minimizes a cost function based on distance and directional difference between adjacent segments as well as the similarity to the device path extracted from the previous frame. The quantitative evaluation was performed using 7 data sets acquired from a canine study. Device shapes with different complexities are compared to semi-automatic segmentations. An average segmentation accuracy of 0.50 0.41 mm was achieved where each point along the device was compared to the point on the reference device centerline with the same distance to the device tip. In all cases the device path could be resolved correctly, which would allow a more accurate and reliable reconstruction of the 3D path of the device.

Accurate segmentation of abdominal organs from medical images is an essential part of surgical planning and computer-aided disease diagnosis. Many existing algorithms are specialized for the segmentation of healthy organs. Cystic pancreas segmentation is especially challenging due to its low contrast boundaries, variability in shape, location and the stage of the pancreatic cancer. We present a semi-automatic segmentation algorithm for pancreata with cysts. In contrast to existing automatic segmentation approaches for healthy pancreas segmentation which are amenable to atlas/statistical shape approaches, a pancreas with cysts can have even higher variability with respect to the shape of the pancreas due to the size and shape of the cyst(s). Hence, fine results are better attained with semi-automatic steerable approaches. We use a novel combination of random walker and region growing approaches to delineate the boundaries of the pancreas and cysts with respective best Dice coefficients of 85.1% and 86.7%, and respective best volumetric overlap errors of 26.0% and 23.5%. Results show that the proposed algorithm for pancreas and pancreatic cyst segmentation is accurate and stable.

In this paper we propose a novel approach towards fast multi-class volume segmentation that exploits supervoxels in order to reduce complexity, time and memory requirements. Current methods for biomedical image segmentation typically require either complex mathematical models with slow convergence, or expensive-to-calculate image features, which makes them non-feasible for large volumes with many objects (tens to hundreds) of different classes, as is typical in modern medical and biological datasets. Recently, graphical models such as Markov Random Fields (MRF) or Conditional Random Fields (CRF) are having a huge impact in different computer vision areas (e.g. image parsing, object detection, object recognition) as they provide global regularization for multiclass problems over an energy minimization framework. These models have yet to find impact in biomedical imaging due to complexities in training and slow inference in 3D images due to the very large number of voxels. Here, we define an interactive segmentation approach over a supervoxel space by first defining novel, robust and fast regional descriptors for supervoxels. Then, a hierarchical segmentation approach is adopted by training Contextual Extremely Random Forests in a user-defined label hierarchy where the classification output of the previous layer is used as additional features to train a new classifier to refine more detailed label information. This hierarchical model yields final class likelihoods for supervoxels which are finally refined by a MRF model for 3D segmentation. Results demonstrate the effectiveness on a challenging cryo-soft X-ray tomography dataset by segmenting cell areas with only a few user scribbles as the input for our algorithm. Further results demonstrate the effectiveness of our method to fully extract different organelles from the cell volume with another few seconds of user interaction.

Despite several decades of research into segmentation techniques, automated medical image segmentation is barely usable in a clinical context, and still at vast user time expense. This paper illustrates unsupervised organ segmentation through the use of a novel automated labelling approximation algorithm followed by a hypersurface front propagation method. The approximation stage relies on a pre-computed image partition forest obtained directly from CT scan data. We have implemented all procedures to operate directly on 3D volumes, rather than slice-by-slice, because our algorithms are dimensionality-independent. The results picture segmentations which identify kidneys, but can easily be extrapolated to other body parts. Quantitative analysis of our automated segmentation compared against hand-segmented gold standards indicates an average Dice similarity coefficient of 90%. Results were obtained over volumes of CT data with 9 kidneys, computing both volume-based similarity measures (such as the Dice and Jaccard coefficients, true positive volume fraction) and size-based measures (such as the relative volume difference). The analysis considered both healthy and diseased kidneys, although extreme pathological cases were excluded from the overall count. Such cases are difficult to segment both manually and automatically due to the large amplitude of Hounsfield unit distribution in the scan, and the wide spread of the tumorous tissue inside the abdomen. In the case of kidneys that have maintained their shape, the similarity range lies around the values obtained for inter-operator variability. Whilst the procedure is fully automated, our tools also provide a light level of manual editing.

We propose a 3D prostate segmentation method for transrectal ultrasound (TRUS) images, which is based on patch-based feature learning framework. Patient-specific anatomical features are extracted from aligned training images and adopted as signatures for each voxel. The most robust and informative features are identified by the feature selection process to train the kernel support vector machine (KSVM). The well-trained SVM was used to localize the prostate of the new patient. Our segmentation technique was validated with a clinical study of 10 patients. The accuracy of our approach was assessed using the manual segmentations (gold standard). The mean volume Dice overlap coefficient was 89.7%. In this study, we have developed a new prostate segmentation approach based on the optimal feature learning framework, demonstrated its clinical feasibility, and validated its accuracy with manual segmentations.

This paper presents a novel airway segmentation method based on intensity structure analysis and graph-cut. Airway segmentation is an important step in analyzing chest CT volumes for computerized lung cancer detection, emphysema diagnosis, asthma diagnosis, and pre- and intra-operative bronchoscope navigation. However, obtaining a complete 3-D airway tree structure from a CT volume is quite challenging. Several researchers have proposed automated algorithms basically based on region growing and machine learning techniques. However these methods failed to detect the peripheral bronchi branches. They caused a large amount of leakage. This paper presents a novel approach that permits more accurate extraction of complex bronchial airway region. Our method are composed of three steps. First, the Hessian analysis is utilized for enhancing the line-like structure in CT volumes, then a multiscale cavity-enhancement filter is employed to detect the cavity-like structure from the previous enhanced result. In the second step, we utilize the support vector machine (SVM) to construct a classifier for removing the FP regions generated. Finally, the graph-cut algorithm is utilized to connect all of the candidate voxels to form an integrated airway tree. We applied this method to sixteen cases of 3D chest CT volumes. The results showed that the branch detection rate of this method can reach about 77.7% without leaking into the lung parenchyma areas.

For patient-specific voxel-based visuo-haptic rendering of CT scans of the liver area, the fully automatic segmentation of large volume structures such as skin, soft tissue, lungs and intestine (risk structures) is important. Using a machine learning based approach, several existing segmentations from 10 segmented gold-standard patients are learned by random decision forests individually and collectively. The core of this paper is feature selection and the application of the learned classifiers to a new patient data set. In a leave-some-out cross-validation, the obtained full volume segmentations are compared to the gold-standard segmentations of the untrained patients. The proposed classifiers use a multi-dimensional feature space to estimate the hidden truth, instead of relying on clinical standard threshold and connectivity based methods. The result of our efficient whole-body section classification are multi-label maps with the considered tissues. For visuo-haptic simulation, other small volume structures would have to be segmented additionally. We also take a look into these structures (liver vessels). For an experimental leave-some-out study consisting of 10 patients, the proposed method performs much more efficiently compared to state of the art methods. In two variants of leave-some-out experiments we obtain best mean DICE ratios of 0.79, 0.97, 0.63 and 0.83 for skin, soft tissue, hard bone and risk structures. Liver structures are segmented with DICE 0.93 for the liver, 0.43 for blood vessels and 0.39 for bile vessels.

Prostate segmentation on 3D MR images is a challenging task due to image artifacts, large inter-patient prostate shape and texture variability, and lack of a clear prostate boundary specifically at apex and base levels. We propose a supervised machine learning model that combines atlas based Active Appearance Model (AAM) with a Deep Learning model to segment the prostate on MR images. The performance of the segmentation method is evaluated on 20 unseen MR image datasets. The proposed method combining AAM and Deep Learning achieves a mean Dice Similarity Coefficient (DSC) of 0.925 for whole 3D MR images of the prostate using axial cross-sections. The proposed model utilizes the adaptive atlas-based AAM model and Deep Learning to achieve significant segmentation accuracy.

Magnetic Resonance Imaging (MRI) is a vital tool in the diagnosis and characterization of multiple sclerosis (MS). MS lesions can be imaged with relatively high contrast using either Fluid Attenuated Inversion Recovery (FLAIR) or Double Inversion Recovery (DIR). Automated segmentation and accurate tracking of MS lesions from MRI remains a challenging problem. Here, an information theoretic approach to cluster the voxels in pseudo-colorized multispectral MR data (FLAIR, DIR, T₂-weighted) is utilized to automatically segment MS lesions of various sizes and noise levels. The Improved Jump Method (IJM) clustering, assisted by edge suppression, is applied to the segmentation of white matter (WM), gray matter (GM), cerebrospinal fluid (CSF) and MS lesions, if present, into a subset of slices determined to be the best MS lesion candidates via Otsu’s method. From this preliminary clustering, the modal data values for the tissues can be determined. A Euclidean distance is then used to estimate the fuzzy memberships of each brain voxel for all tissue types and their 50/50 partial volumes. From these estimates, binary discrete and fuzzy MS lesion masks are constructed. Validation is provided by using three synthetic MS lesions brains (mild, moderate and severe) with labeled ground truths. The MS lesions of mild, moderate and severe designations were detected with a sensitivity of 83.2%, and 88.5%, and 94.5%, and with the corresponding Dice similarity coefficient (DSC) of 0.7098, 0.8739, and 0.8266, respectively. The effect of MRI noise is also examined by simulated noise and the application of a bilateral filter in preprocessing.

Accurate lung vessel segmentation is an important operation for lung CT analysis. Filters that are based on analyzing the eigenvalues of the Hessian matrix are popular for pulmonary vessel enhancement. However, due to their low response at vessel bifurcations and vessel boundaries, extracting lung vessels by thresholding the vesselness is not sufficiently accurate. Some methods turn to graph-cuts for more accurate segmentation, as it incorporates neighbourhood information. In this work, we propose a new graph-cuts cost function combining appearance and shape, where CT intensity represents appearance and vesselness from a Hessian-based filter represents shape. Due to the amount of voxels in high resolution CT scans, the memory requirement and time consumption for building a graph structure is very high. In order to make the graph representation computationally tractable, those voxels that are considered clearly background are removed from the graph nodes, using a threshold on the vesselness map. The graph structure is then established based on the remaining voxel nodes, source/sink nodes and the neighbourhood relationship of the remaining voxels. Vessels are segmented by minimizing the energy cost function with the graph-cuts optimization framework. We optimized the parameters used in the graph-cuts cost function and evaluated the proposed method with two manually labeled sub-volumes. For independent evaluation, we used 20 CT scans of the VESSEL12 challenge. The evaluation results of the sub-volume data show that the proposed method produced a more accurate vessel segmentation compared to the previous methods, with F1 score 0.76 and 0.69. In the VESSEL12 data-set, our method obtained a competitive performance with an area under the ROC curve of 0.975, especially among the binary submissions.

Inspired by the diffusion of a particle, we present a novel approach for performing a semiautomatic segmentation of tomographic images in 3D, 4D or higher dimensions to meet the requirements of high-throughput measurements in a synchrotron X-ray microtomograph. Given a small number of 2D-slices with at least two manually labeled segments, one can either analytically determine the probability that an intelligently weighted random walk starting at one labeled pixel will be at a certain time at a specific position in the dataset or determine the probability approximately by performing several random walks. While the weights of a random walk take into account local information at the starting point, the random walk itself can be in any dimension. Starting a great number of random walks in each labeled pixel, a voxel in the dataset will be hit by several random walks over time. Hence, the image can be segmented by assigning each voxel to the label where the random walks most likely started from. Due to the high scalability of random walks, this approach is suitable for high throughput measurements. Additionally, we describe an interactively adjusted active contours slice by slice method considering local information, where we start with one manually labeled slice and move forward in any direction. This approach is superior with respect to accuracy towards the diffusion algorithm but inferior in the amount of tedious manual processing steps. The methods were applied on 3D and 4D datasets and evaluated by means of manually labeled images obtained in a realistic scenario with biologists.

This paper presents a novel renal artery segmentation method combining graph-cut and template-based tracking methods and its application to estimation of renal vascular dominant region. For the purpose of giving a computer assisted diagnose for kidney surgery planning, it is important to obtain the correct topological structures of renal artery for estimation of renal vascular dominant regions. Renal artery has a low contrast, and its precise extraction is a difficult task. Previous method utilizing vesselness measure based on Hessian analysis, still cannot extract the tiny blood vessels in low-contrast area. Although model-based methods including superellipsoid model or cylindrical intensity model are low-contrast sensitive to the tiny blood vessels, problems including over-segmentation and poor bifurcations detection still remain. In this paper, we propose a novel blood vessel segmentation method combining a new Hessian-based graph-cut and template modeling tracking method. Firstly, graph-cut algorithm is utilized to obtain the rough segmentation result. Then template model tracking method is utilized to improve the accuracy of tiny blood vessel segmentation result. Rough segmentation utilizing graph-cut solves the bifurcations detection problem effectively. Precise segmentation utilizing template model tracking focuses on the segmentation of tiny blood vessels. By combining these two approaches, our proposed method segmented 70% of the renal artery of 1mm in diameter or larger. In addition, we demonstrate such precise segmentation can contribute to divide renal regions into a set of blood vessel dominant regions utilizing Voronoi diagram method.

Purpose: Magnetic resonance imaging is nowadays the hallmark to diagnose multiple sclerosis (MS), characterized by white matter lesions. Several approaches have been recently presented to tackle the lesion segmentation problem, but none of them have been accepted as a standard tool in the daily clinical practice. In this work we present yet another tool able to automatically segment white matter lesions outperforming the current-state-of-the-art approaches. Methods: This work is an extension of Roura et al. [1], where external and platform dependent pre-processing libraries (brain extraction, noise reduction and intensity normalization) were required to achieve an optimal performance. Here we have updated and included all these required pre-processing steps into a single framework (SPM software). Therefore, there is no need of external tools to achieve the desired segmentation results. Besides, we have changed the working space from T1w to FLAIR, reducing interpolation errors produced in the registration process from FLAIR to T1w space. Finally a post-processing constraint based on shape and location has been added to reduce false positive detections. Results: The evaluation of the tool has been done on 24 MS patients. Qualitative and quantitative results are shown with both approaches in terms of lesion detection and segmentation. Conclusion: We have simplified both installation and implementation of the approach, providing a multiplatform tool1 integrated into the SPM software, which relies only on using T1w and FLAIR images. We have reduced with this new version the computation time of the previous approach while maintaining the performance.

Computed tomography (CT) imaging is a sensitive and specific lung cancer screening tool for the high-risk population and shown to be promising for detection of lung cancer. This study proposes an automatic methodology for detecting and segmenting lung nodules from CT images. The proposed methods begin with thorax segmentation, lung extraction and reconstruction of the original shape of the parenchyma using morphology operations. Next, a multi-scale hessian-based vesselness filter is applied to extract lung vasculature in lung. The lung vasculature mask is subtracted from the lung region segmentation mask to extract 3D regions representing candidate pulmonary nodules. Finally, the remaining structures are classified as nodules through shape and intensity features which are together used to train an artificial neural network. Up to 75% sensitivity and 98% specificity was achieved for detection of lung nodules in our testing dataset, with an overall accuracy of 97.62%±0.72% using 11 selected features as input to the neural network classifier, based on 4-fold cross-validation studies. Receiver operator characteristics for identifying nodules revealed an area under curve of 0.9476.

Optical coherence tomography (OCT) is used to produce high-resolution three-dimensional images of the retina, which permit the investigation of retinal irregularities. In dry age-related macular degeneration (AMD), a chronic eye disease that causes central vision loss, disruptions such as drusen and changes in retinal layer thicknesses occur which could be used as biomarkers for disease monitoring and diagnosis. Due to the topology disrupting pathology, existing segmentation methods often fail. Here, we present a solution for the segmentation of retinal layers in dry AMD subjects by extending our previously presented loosely coupled level sets framework which operates on attenuation coefficients. In eyes affected by AMD, Bruch’s membrane becomes visible only below the drusen and our segmentation framework is adapted to delineate such a partially discernible interface. Furthermore, the initialization stage, which tentatively segments five interfaces, is modified to accommodate the appearance of drusen. This stage is based on Dijkstra's algorithm and combines prior knowledge on the shape of the interface, gradient and attenuation coefficient in the newly proposed cost function. This prior knowledge is incorporated by varying the weights for horizontal, diagonal and vertical edges. Finally, quantitative evaluation of the accuracy shows a good agreement between manual and automated segmentation.

In multi-atlas segmentation, one typically registers several atlases to the new image, and their respective segmented label images are transformed and fused to form the final segmentation. After each registration, the quality of the registration is reflected by the single global value: the final registration cost. Ideally, if the quality of the registration can be evaluated at each point, independent of the registration process, which also provides a direction in which the deformation can further be improved, the overall segmentation performance can be improved. We propose such a self-correcting multi-atlas segmentation method. The method is applied on hippocampus segmentation from brain images and statistically significantly improvement is observed.

The cerebellum is a somatotopically organized central component of the central nervous system well known to be involved with motor coordination and increasingly recognized roles in cognition and planning. Recent work in multi-atlas labeling has created methods that offer the potential for fully automated 3-D parcellation of the cerebellar lobules and vermis (which are organizationally equivalent to cortical gray matter areas). This work explores the trade offs of using different statistical fusion techniques and post hoc optimizations in two datasets with distinct imaging protocols. We offer a novel fusion technique by extending the ideas of the Selective and Iterative Method for Performance Level Estimation (SIMPLE) to a patch-based performance model. We demonstrate the effectiveness of our algorithm, Non-Local SIMPLE, for segmentation of a mixed population of healthy subjects and patients with severe cerebellar anatomy. Under the first imaging protocol, we show that Non-Local SIMPLE outperforms previous gold-standard segmentation techniques. In the second imaging protocol, we show that Non-Local SIMPLE outperforms previous gold standard techniques but is outperformed by a non-locally weighted vote with the deeper population of atlases available. This work advances the state of the art in open source cerebellar segmentation algorithms and offers the opportunity for routinely including cerebellar segmentation in magnetic resonance imaging studies that acquire whole brain T1-weighted volumes with approximately 1 mm isotropic resolution.

We propose a fully-automatic method for brain tumor segmentation that does not require any training phase. Our approach is based on a sequence of segmentations using the Mumford-Shah cartoon model with varying parameters. In order to come up with a very fast implementation, we extend the recent primal-dual algorithm of Strekalovskiy et al. (2014) from the 2D to the medically relevant 3D setting. Moreover, we suggest a new confidence refinement and show that it can increase the precision of our segmentations substantially. Our method is evaluated on 188 data sets with high-grade gliomas and 25 with low-grade gliomas from the BraTS14 database. Within a computation time of only three minutes, we achieve Dice scores that are comparable to state-of-the-art methods.

Image segmentation is an important problem in computer vision and has significant applications in the segmentation of cellular images. Many different imaging techniques exist and produce a variety of image properties which pose difficulties to image segmentation routines. Bright-field images are particularly challenging because of the non-uniform shape of the cells, the low contrast between cells and background, and imaging artifacts such as halos and broken edges. Classical segmentation techniques often produce poor results on these challenging images. Previous attempts at bright-field imaging are often limited in scope to the images that they segment. In this paper, we introduce a new algorithm for automatically segmenting cellular images. The algorithm incorporates two game theoretic models which allow each pixel to act as an independent agent with the goal of selecting their best labelling strategy. In the non-cooperative model, the pixels choose strategies greedily based only on local information. In the cooperative model, the pixels can form coalitions, which select labelling strategies that benefit the entire group. Combining these two models produces a method which allows the pixels to balance both local and global information when selecting their label. With the addition of k-means and active contour techniques for initialization and post-processing purposes, we achieve a robust segmentation routine. The algorithm is applied to several cell image datasets including bright-field images, fluorescent images and simulated images. Experiments show that the algorithm produces good segmentation results across the variety of datasets which differ in cell density, cell shape, contrast, and noise levels.

Cochlear implants can restore hearing to deaf or partially deaf patients. In order to plan the intervention, a model from high resolution µCT images is to be built from accurate cochlea segmentations and then, adapted to a patient-specific model. Thus, a precise segmentation is required to build such a model. We propose a new framework for segmentation of µCT cochlear images using random walks where a region term is combined with a distance shape prior weighted by a confidence map to adjust its influence according to the strength of the image contour. Then, the region term can take advantage of the high contrast between the background and foreground and the distance prior guides the segmentation to the exterior of the cochlea as well as to less contrasted regions inside the cochlea. Finally, a refinement is performed preserving the topology using a topological method and an error control map to prevent boundary leakage. We tested the proposed approach with 10 datasets and compared it with the latest techniques with random walks and priors. The experiments suggest that this method gives promising results for cochlea segmentation.

Positron Emission Tomography (PET) and Computed Tomography (CT) have been widely used in clinical practice for radiation therapy. Most existing methods only used one image modality, either PET or CT, which suffers from the low spatial resolution in PET or low contrast in CT. In this paper, a novel 3D graph cut method is proposed, which integrated Gaussian Mixture Models (GMMs) into the graph cut method. We also employed the random walk method as an initialization step to provide object seeds for the improvement of the graph cut based segmentation on PET and CT images. The constructed graph consists of two sub-graphs and a special link between the sub-graphs which penalize the difference segmentation between the two modalities. Finally, the segmentation problem is solved by the max-flow/min-cut method. The proposed method was tested on 20 patients’ PET-CT images, and the experimental results demonstrated the accuracy and efficiency of the proposed algorithm.

In this paper, a fully automatic method is proposed to segment the lung tumor in clinical 3D PET-CT images. The proposed method effectively combines PET and CT information to make full use of the high contrast of PET images and superior spatial resolution of CT images. Our approach consists of three main parts: (1) initial segmentation, in which spines are removed in CT images and initial connected regions achieved by thresholding based segmentation in PET images; (2) coarse segmentation, in which monotonic downhill function is applied to rule out structures which have similar standardized uptake values (SUV) to the lung tumor but do not satisfy a monotonic property in PET images; (3) fine segmentation, random forests method is applied to accurately segment the lung tumor by extracting effective features from PET and CT images simultaneously. We validated our algorithm on a dataset which consists of 24 3D PET-CT images from different patients with non-small cell lung cancer (NSCLC). The average TPVF, FPVF and accuracy rate (ACC) were 83.65%, 0.05% and 99.93%, respectively. The correlation analysis shows our segmented lung tumor volumes has strong correlation ( average 0.985) with the ground truth 1 and ground truth 2 labeled by a clinical expert.

Probabilistic atlas based on human anatomical structure has been widely used for organ segmentation. The challenge is how to register the probabilistic atlas to the patient volume. Additionally, there is the disadvantage that the conventional probabilistic atlas may cause a bias toward the specific patient study due to a single reference. Hence, we propose a template matching framework based on an iterative probabilistic atlas for organ segmentation. Firstly, we find a bounding box for the organ based on human anatomical localization. Then, the probabilistic atlas is used as a template to find the organ in this bounding box by using template matching technology. Comparing our method with conventional and recently developed atlas-based methods, our results show an improvement in the segmentation accuracy for multiple organs (p < 0:00001).

Automated detection and segmentation of vertebral bodies from spinal computed tomography (CT) images is usually a prerequisite step for numerous spine-related medical applications, such as diagnosis, surgical planning and follow-up assessment of spinal pathologies. However, automated detection and segmentation are challenging tasks due to a relatively high degree of anatomical complexity, presence of unclear boundaries and articulation of vertebrae with each other. In this paper, we describe a sparse representation error minimization (SEM) framework for joint detection and segmentation of vertebral bodies in CT images. By minimizing the sparse representation error of sampled intensity values, we are able to recover the oriented bounding box (OBB) and segmentation binary mask for each vertebral body in the CT image. The performance of the proposed SEM framework was evaluated on five CT images of the thoracolumbar spine. The resulting Euclidean distance of 1:75±1:02 mm, computed between the center points of recovered and corresponding reference OBBs, and Dice coefficient of 92:3±2:7%, computed between the resulting and corresponding reference segmentation binary masks, indicate that the proposed framework can successfully detect and segment vertebral bodies in CT images of the thoracolumbar spine.

This paper proposes a new method for mapping volume models of human tissues onto a target volume with simple shapes. The proposed method is based on our modified self-organizing deformable model (mSDM)^{1, 2} which finds the one-to-one mapping with no foldovers between arbitrary surface model and a target surface. By extending mSDM to apply to volume models, the proposed method, called volumetric SDM (vSDM), establishes the one-to-one correspondence between the tissue volume model and its target volume. At the same time, vSDM can preserve geometrical properties of the original model before and after mapping. This characteristic of vSDM makes it easy to find the correspondence between tissue models.

We present a novel approach to parcellate – delineate the anatomical feature (folds, gyri, sulci) boundaries – the brain cortex. Our approach is based on extracting the 3D brain cortical surface mesh from magnetic resonance (MR) images, computing the shape measures (area, mean curvature, geodesic, and travel depths) for this mesh, and delineating the anatomical feature boundaries using these measures. We use angle-area preserving mapping of the cortical surface mesh to a simpler topology (disk or rectangle) to aid in the visualization and delineation of these boundaries. Contrary to commonly used generic 2D brain image atlas-based approaches, we use 3D surface mesh data extracted from a given brain MR imaging data and its specific shape measures for the parcellation. Our method does not require any non-linear registration of a given brain dataset to a generic atlas and hence, does away with the structure similarity assumption critical to the atlas-based approaches. We evaluate our approach using Mindboggle manually labeled brain datasets and achieve the following accuracies: 72.4% for gyri, 78.5% for major sulci, and 98.4% for folds. These results warrant further investigation of this approach as an alternative or as an initialization to the atlas-based approaches.

Aortic valve (AV) disease is a significant cause of morbidity and mortality. The preferred treatment modality for severe AV disease is surgical resection and replacement of the native valve with either a mechanical or tissue prosthetic. In order to develop effective and long-lasting treatment methods, computational analyses, e.g., structural finite element (FE) and computational fluid dynamic simulations, are very effective for studying valve biomechanics. These computational analyses are based on mesh models of the aortic valve, which are usually constructed from 3D CT images though many hours of manual annotation, and therefore an automatic valve shape reconstruction method is desired. In this paper, we present a method for estimating the aortic valve shape from 3D cardiac CT images, which is represented by triangle meshes. We propose a pipeline for aortic valve shape estimation which includes novel algorithms for building local shape dictionaries and for building landmark detectors and curve detectors using local shape dictionaries. The method is evaluated on real patient image dataset using a leave-one-out approach and achieves an average accuracy of 0.69 mm. The work will facilitate automatic patient-specific computational modeling of the aortic valve.

Statistical shape and appearance models are often based on the accurate identification of one-to-one correspondences in a training data set. At the same time, the determination of these corresponding landmarks is the most challenging part of such methods. Hufnagel et al.¹ developed an alternative method using correspondence probabilities for a statistical shape model. In Krüuger et al.^{2, 3} we propose the use of probabilistic correspondences for statistical appearance models by incorporating appearance information into the framework. We employ a point-based representation of image data combining position and appearance information. The model is optimized and adapted by a maximum a-posteriori (MAP) approach deriving a single global optimization criterion with respect to model parameters and observation dependent parameters that directly affects shape and appearance information of the considered structures. Because initially unknown correspondence probabilities are used and a higher number of degrees of freedom is introduced to the model a regularization of the model generation process is advantageous. For this purpose we extend the derived global criterion by a regularization term which penalizes implausible topological changes. Furthermore, we propose a multi-level approach for the optimization, to increase the robustness of the model generation process.

Optimization-based segmentation approaches deriving from discrete graph-cuts and continuous max-flow have become increasingly nuanced, allowing for topological and geometric constraints on the resulting segmentation while retaining global optimality. However, these two considerations, topological and geometric, have yet to be combined in a unified manner. This paper presents the concept of shape complexes, which combine geodesic star convexity with extendable continuous max-flow solvers. These shape complexes allow more complicated shapes to be created through the use of multiple labels and super-labels, with geodesic star convexity governed by a topological ordering. These problems can be optimized using extendable continuous max-flow solvers. Previous work required computationally expensive co-ordinate system warping which are ill-defined and ambiguous in the general case. These shape complexes are validated in a set of synthetic images as well as atrial wall segmentation from contrast-enhanced CT. Shape complexes represent a new, extendable tool alongside other continuous max-flow methods that may be suitable for a wide range of medical image segmentation problems.

An accurate ventricular function quantification is important to support evaluation, diagnosis and prognosis of several cardiac pathologies. However, expert heart delineation, specifically for the right ventricle, is a time consuming task with high inter-and-intra observer variability. A fully automatic 3D+time heart segmentation framework is herein proposed for short-axis-cardiac MRI sequences. This approach estimates the heart using exclusively information from the sequence itself without tuning any parameters. The proposed framework uses a coarse-to-fine approach, which starts by localizing the heart via spatio-temporal analysis, followed by a segmentation of the basal heart that is then propagated to the apex by using a non-rigid-registration strategy. The obtained volume is then refined by estimating the ventricular muscle by locally searching a prior endocardium- pericardium intensity pattern. The proposed framework was applied to 48 patients datasets supplied by the organizers of the MICCAI 2012 Right Ventricle segmentation challenge. Results show the robustness, efficiency and competitiveness of the proposed method both in terms of accuracy and computational load.

Among all types of cancer, gynecological malignancies belong to the 4th most frequent type of cancer among women. Besides chemotherapy and external beam radiation, brachytherapy is the standard procedure for the treatment of these malignancies. In the progress of treatment planning, localization of the tumor as the target volume and adjacent organs of risks by segmentation is crucial to accomplish an optimal radiation distribution to the tumor while simultaneously preserving healthy tissue. Segmentation is performed manually and represents a time-consuming task in clinical daily routine. This study focuses on the segmentation of the rectum/sigmoid colon as an Organ-At-Risk in gynecological brachytherapy. The proposed segmentation method uses an interactive, graph-based segmentation scheme with a user-defined template. The scheme creates a directed two dimensional graph, followed by the minimal cost closed set computation on the graph, resulting in an outlining of the rectum. The graphs outline is dynamically adapted to the last calculated cut. Evaluation was performed by comparing manual segmentations of the rectum/sigmoid colon to results achieved with the proposed method. The comparison of the algorithmic to manual results yielded to a Dice Similarity Coefficient value of 83.85±4.08%, in comparison to 83.97±8.08% for the comparison of two manual segmentations of the same physician. Utilizing the proposed methodology resulted in a median time of 128 seconds per dataset, compared to 300 seconds needed for pure manual segmentation.

Breast cancer is a one of the most common forms of cancer in terms of new cases and deaths both in the United States and worldwide. However, the survival rate with breast cancer is high if it is detected and treated before it spreads to other parts of the body. The most common screening methods for breast cancer are mammography and digital tomosynthesis, which involve acquiring X-ray images of the breasts that are interpreted by radiologists. The work described in this paper is aimed at optimizing the presentation of mammography and tomosynthesis images to the radiologist, thereby improving the early detection rate of breast cancer and the resulting patient outcomes. Breast cancer tissue has greater density than normal breast tissue, and appears as dense white image regions that are asymmetrical between the breasts. These irregularities are easily seen if the breast images are aligned and viewed side-by-side. However, since the breasts are imaged separately during mammography, the images may be poorly centered and aligned relative to each other, and may not properly focus on the tissue area. Similarly, although a full three dimensional reconstruction can be created from digital tomosynthesis images, the same centering and alignment issues can occur for digital tomosynthesis. Thus, a preprocessing algorithm that aligns the breasts for easy side-by-side comparison has the potential to greatly increase the speed and accuracy of mammogram reading. Likewise, the same preprocessing can improve the results of automatic tissue classification algorithms for mammography. In this paper, we present an automated segmentation algorithm for mammogram and tomosynthesis images that aims to improve the speed and accuracy of breast cancer screening by mitigating the above mentioned problems. Our algorithm uses information in the DICOM header to facilitate preprocessing, and incorporates anatomical region segmentation and contour analysis, along with a hidden Markov model (HMM) for processing the multi-frame tomosynthesis images. The output of the algorithm is a new set of images that have been processed to show only the diagnostically relevant region and align the breasts so that they can be easily compared side-by-side. Our method has been tested on approximately 750 images, including various examples of mammogram, tomosynthesis, and scanned images, and has correctly segmented the diagnostically relevant image region in 97% of cases.

This paper deals with separation of merged Langerhans islets in segmentations in order to evaluate correct histogram of islet diameters. A distribution of islet diameters is useful for determining the feasibility of islet transplantation in diabetes. First, the merged islets at training segmentations are manually separated by medical experts. Based on the single islets, the merged islets are identified and the SVM classifier is trained on both classes (merged/single islets). The testing segmentations were over-segmented using watershed transform and the most probable back merging of islets were found using trained SVM classifier. Finally, the optimized segmentation is compared with ground truth segmentation (correctly separated islets).

Spinal deformity correction with vertebral fixation is nowadays the preferred surgical treatment, where pedicle screws are inserted through pedicles into corresponding vertebral bodies and afterwards connected with rods. In clinical practice, the straight-forward and anatomical insertion technique are currently being used for pedicle screw placement surgery. However, it is difficult to quantitatively compare both techniques and determine which technique is more adequate for each planned pedicle screw before surgery (i.e. preoperatively). In this paper, we therefore describe a framework for quantitative comparison between the straight-forward and anatomical insertion technique for pedicle screw placement surgery by evaluating the screw fastening strength. Quantitative comparisons were performed on computed tomography images of 11 patients with 74 manually planned pedicle screws, who underwent the vertebral fixation procedure. The first quantitative comparison was performed between the straight-forward and anatomical pedicle screw insertion technique, which resulted in a relatively high agreement with mean absolute difference of 0.0mm in screw diameter, 2.9mm in screw length, 1.2mm in pedicle crossing point and 6.5° in screw inclinations. The second quantitative comparison was performed between the best resulting pedicle screw insertion technique and manually obtained pedicle screw plans, which again resulted in a relatively high agreement with mean absolute difference of 0.5mm in screw diameter, 4.7mm in screw length, 2.4mm in pedicle crossing point and 6.0° in screw inclinations. Both the straight-forward and anatomical insertion technique proved approximately equal in terms of the screw fastening strength.

Three-dimensional reconstructions of coronary arterial trees from X-ray rotational angiography (RA) images have the potential to compensate the limitations of RA due to projective imaging. Most of the existing model based reconstruction algorithms are either based on forward-projection of a 3D deformable model onto X-ray angiography images or back-projection of 2D information extracted from X-ray angiography images to 3D space for further processing. All of these methods have their shortcomings such as dependency on accurate 2D centerline segmentations. In this paper, the reconstruction is approached from a novel perspective, and is formulated as a probabilistic reconstruction method based on mixture model (MM) representation of point sets describing the coronary arteries. Specifically, it is assumed that the coronary arteries could be represented by a set of 3D points, whose spatial locations denote the Gaussian components in the MM. Additionally, an extra uniform distribution is incorporated in the mixture model to accommodate outliers (noise, over-segmentation etc.) in the 2D centerline segmentations. Treating the given 2D centreline segmentations as data points generated from MM, the 3D means, isotropic variance, and mixture weights of the Gaussian components are estimated by maximizing a likelihood function. Initial results from a phantom study show that the proposed method is able to handle outliers in 2D centreline segmentations, which indicates the potential of our formulation. Preliminary reconstruction results in the clinical data are also presented.

We present a simple approach to the voxelwise classification of brain tissue acquired with diffusion weighted MRI (DWI). The approach leverages the power of spherical harmonics to summarise the diffusion information, sampled at many points over a sphere, using only a handful of coefficients. We use simple features that are invariant to the rotation of the highly orientational diffusion data. This provides a way to directly classify voxels whose diffusion characteristics are similar yet whose primary diffusion orientations differ. Subsequent application of machine-learning to the spherical harmonic coefficients therefore may permit classification of DWI voxels according to their inferred underlying fibre properties, whilst ignoring the specifics of orientation. After smoothing apparent diffusion coefficients volumes, we apply a spherical harmonic transform, which models the multi-directional diffusion data as a collection of spherical basis functions. We use the derived coefficients as voxelwise feature vectors for classification. Using a simple Gaussian mixture model, we examined the classification performance for a range of sub-classes (3-20). The results were compared against existing alternatives for tissue classification e.g. fractional anisotropy (FA) or the standard model used by Camino.1 The approach was implemented on both two publicly-available datasets: an ex-vivo pig brain and in-vivo human brain from the Human Connectome Project (HCP). We have demonstrated how a robust classification of DWI data can be performed without the need for a model reconstruction step. This avoids the potential confounds and uncertainty that such models may impose, and has the benefit of being computable directly from the DWI volumes. As such, the method could prove useful in subsequent pre-processing stages, such as model fitting, where it could inform about individual voxel complexities and improve model parameter choice.

Echo-planar imaging (EPI) is commonly used for diffusion-weighted imaging (DWI) but is susceptible to nonlinear geometric distortions arising from inhomogeneities in the static magnetic field. These inhomogeneities can be measured and corrected using a fieldmap image acquired during the scanning process. In studies where the fieldmap image is not collected, these distortions can be corrected, to some extent, by nonlinearly registering the diffusion image to a corresponding anatomical image, either a T1- or T2-weighted image. Here we compared two EPI distortion correction pipelines, both based on nonlinear registration, which were optimized for the particular weighting of the structural image registration target. The first pipeline used a 3D nonlinear registration to a T1-weighted target, while the second pipeline used a 1D nonlinear registration to a T2-weighted target. We assessed each pipeline in its ability to characterize high-level measures of brain connectivity in Parkinson’s disease (PD) in 189 individuals (58 healthy controls, 131 people with PD) from the Parkinson’s Progression Markers Initiative (PPMI) dataset. We computed a structural connectome (connectivity map) for each participant using regions of interest from a cortical parcellation combined with DWI-based whole-brain tractography. We evaluated test-retest reliability of the connectome for each EPI distortion correction pipeline using a second diffusion scan acquired directly after the participants’ first. Finally, we used support vector machine (SVM) classification to assess how accurately each pipeline classified PD versus healthy controls using each participants’ structural connectome.

Regional analysis is normally done by fitting models per voxel and then averaging over a region, accounting for partial volume (PV) only to some degree. In thin, folded regions such as the cerebral cortex, such methods do not work well, as the partial volume confounds parameter estimation. Instead, we propose to fit the models per region directly with explicit PV modeling. In this work we robustly estimate region-wise parameters whilst explicitly accounting for partial volume effects. We use a high-resolution segmentation from a T1 scan to assign each voxel in the diffusion image a probabilistic membership to each of k tissue classes. We rotate the DW signal at each voxel so that it aligns with the z-axis, then model the signal at each voxel as a linear superposition of a representative signal from each of the k tissue types. Fitting involves optimising these representative signals to best match the data, given the known probabilities of belonging to each tissue type that we obtained from the segmentation. We demonstrate this method improves parameter estimation in digital phantoms for the diffusion tensor (DT) and ‘Neurite Orientation Dispersion and Density Imaging’ (NODDI) models. The method provides accurate parameter estimates even in regions where the normal approach fails completely, for example where partial volume is present in every voxel. Finally, we apply this model to brain data from preterm infants, where the thin, convoluted, maturing cortex necessitates such an approach.

Subject motion in a large number of diffusion weighted images (DWIs) for q-space analysis was detected and corrected by using a simple protocol to add multiple interleaved b0 images between each DWI set and at the very end of data acquisition. The realignment matrix was determined from each b0 image with respect to the first b0 image and the matrix was used to realign not only the b0 image itself but also its subsequent DWI set. Degree of improvement in q-space analysis was estimated by calculating total residual sum of squares (RSS) in bi-exponential curve fitting process and also on the fractional anisotropy (FA) of zero displacement (ZDP). The large RSS regions were considerably diminished by realignment at the edges between cerebral gyri and sulci and at the ventricle boundaries in the original images. The large RSS regions around basal ganglia and near the ventricles were kept even by realignment but considerably suppressed with the averaged b0 image for decay-curve estimation. The volume average of RSS was reduced by the maximum of 77% in four volunteers’ results with both the realignment and the averaged b0 images. The FA-ZDP images revealed the improvement by realignment such as the contrast of corpus callosum and suppression of abnormal FA at cerebral sulcus. The number of additional b0 images accounted for 3% of the total number of DWIs, which suggests its feasibility for future clinical application.

Diffusion-weighted MR imaging (DWI) is a powerful tool to study brain tissue microstructure. DWI is sensitive to subtle changes in the white matter (WM), and can provide insight into abnormal brain changes in diseases such as Alzheimer’s disease (AD). In this study, we used 7-Tesla hybrid diffusion imaging (HYDI) to scan 3 transgenic rats (line TgF344-AD; that model the full clinico-pathological spectrum of the human disease) ex vivo at 10, 15 and 24 months. We acquired 300 DWI volumes across 5 q-sampling shells (b=1000, 3000, 4000, 8000, 12000 s/mm²). From the top three b-value shells with highest signal-to-noise ratios, we reconstructed markers of WM disease, including indices of axon density and diameter in the corpus callosum (CC) – directly quantifying processes that occur in AD. As expected, apparent anisotropy progressively decreased with age; there were also decreases in the intra- and extra-axonal MR signal along axons. Axonal diameters were larger in segments of the CC (splenium and body, but not genu), possibly indicating neuritic dystrophy – characterized by enlarged axons and dendrites as previously observed at the ultrastructural level (see Cohen et al., J. Neurosci. 2013). This was further supported by increases in MR signals trapped in glial cells, CSF and possibly other small compartments in WM structures. Finally, tractography detected fewer fibers in the CC at 10 versus 24 months of age. These novel findings offer great potential to provide technical and scientific insight into the biology of brain disease.

Automatic thalamus segmentation is useful to track changes in thalamic volume over time. In this work, we introduce a task-driven dictionary learning framework to find the optimal dictionary given a set of eleven features obtained from T1-weighted MRI and diffusion tensor imaging. In this dictionary learning framework, a linear classifier is designed concurrently to classify voxels as belonging to the thalamus or non-thalamus class. Morphological post-processing is applied to produce the final thalamus segmentation. Due to the uneven size of the training data samples for the non-thalamus and thalamus classes, a non-uniform sampling scheme is pro- posed to train the classifier to better discriminate between the two classes around the boundary of the thalamus. Experiments are conducted on data collected from 22 subjects with manually delineated ground truth. The experimental results are promising in terms of improvements in the Dice coefficient of the thalamus segmentation overstate-of-the-art atlas-based thalamus segmentation algorithms.

We explore a new approach for structural connectivity based segmentations of subcortical brain regions. Connectivity based segmentations are usually based on fibre connections from a seed region to predefined target regions. We present a method for finding significantly connected voxels based on the distribution of connection strengths. Paths from seed voxels to all voxels in a target region are obtained from a shortest-path tractography. For each seed voxel we approximate the distribution with a histogram of path scores. We hypothesise that the majority of estimated connections are false-positives and that their connection strength is distributed differently from true-positive connections. Therefore, an empirical null-distribution is defined for each target region as the average normalized histogram over all voxels in the seed region. Single histograms are then tested against the corresponding null-distribution and significance is determined using the false discovery rate (FDR). Segmentations are based on significantly connected voxels and their FDR. In this work we focus on the thalamus and the target regions were chosen by dividing the cortex into a prefrontal/temporal zone, motor zone, somatosensory zone and a parieto-occipital zone. The obtained segmentations consistently show a sparse number of significantly connected voxels that are located near the surface of the anterior thalamus over a population of 38 subjects.

Segmentation of the thalamus and thalamic nuclei is useful to quantify volumetric changes from neurodegenerative diseases. Most thalamus segmentation algorithms only use T1-weighted magnetic resonance images and current thalamic parcellation methods require manual interaction. Smaller nuclei, such as the lateral and medial geniculates, are challenging to locate due to their small size. We propose an automated segmentation algorithm using a set of features derived from diffusion tensor image (DTI) and thalamic nuclei location priors. After extracting features, a hierarchical random forest classifier is trained to locate the thalamus. A second random forest classifies thalamus voxels as belonging to one of six thalamic nuclei classes. The proposed algorithm was tested using a leave-one-out cross validation scheme and compared with state-of-the-art algorithms. The proposed algorithm has a higher Dice score compared to other methods for the whole thalamus and several nuclei.

Modern magnetic resonance imaging (MRI) brain atlases are high quality 3-D volumes with specific structures labeled in the volume. Atlases are essential in providing a common space for interpretation of results across studies, for anatomical education, and providing quantitative image-based navigation. Extensive work has been devoted to atlas construction for humans, macaque, and several non-primate species (e.g., rat). One notable gap in the literature is the common squirrel monkey – for which the primary published atlases date from the 1960’s. The common squirrel monkey has been used extensively as surrogate for humans in biomedical studies, given its anatomical neuro-system similarities and practical considerations. This work describes the continued development of a multi-modal MRI atlas for the common squirrel monkey, for which a structural imaging space and gray matter parcels have been previously constructed. This study adds white matter tracts to the atlas. The new atlas includes 49 white matter (WM) tracts, defined using diffusion tensor imaging (DTI) in three animals and combines these data to define the anatomical locations of these tracks in a standardized coordinate system compatible with previous development. An anatomist reviewed the resulting tracts and the inter-animal reproducibility (i.e., the Dice index of each WM parcel across animals in common space) was assessed. The Dice indices range from 0.05 to 0.80 due to differences of local registration quality and the variation of WM tract position across individuals. However, the combined WM labels from the 3 animals represent the general locations of WM parcels, adding basic connectivity information to the atlas.

Diffusion-weighted magnetic resonance imaging (DW-MRI) is a powerful tool to probe the diffusion of water through tissues. Through the application of magnetic gradients of appropriate direction, intensity and duration constituting the acquisition parameters, information can be retrieved about the underlying microstructural organization of the brain. In this context, an important and open question is to determine an optimal sequence of such acquisition parameters for a specific purpose. The use of simulated DW-MRI data for a given microstructural configuration provides a convenient and efficient way to address this problem. We first present a novel hybrid method for the synthetic simulation of DW-MRI signals that combines analytic expressions in simple geometries such as spheres and cylinders and Monte Carlo (MC) simulations elsewhere. Our hybrid method remains valid for any acquisition parameters and provides identical levels of accuracy with a computational time that is 90% shorter than that required by MC simulations for commonly-encountered microstructural configurations. We apply our novel simulation technique to estimate the radius of axons under various noise levels with different acquisition protocols commonly used in the literature. The results of our comparison suggest that protocols favoring a large number of gradient intensities such as a Cube and Sphere (CUSP) imaging provide more accurate radius estimation than conventional single-shell HARDI acquisitions for an identical acquisition time.

Sickle cell disease (SCD) is an inherited blood disorder that effects red blood cells, which can lead to vasoocclusion, ischemia and infarct. This disease often results in neurological damage and strokes, leading to morbidity and mortality. Functional Magnetic Resonance Imaging (fMRI) is a non-invasive technique for measuring and mapping the brain activity. Blood Oxygenation Level-Dependent (BOLD) signals contain also information about the neurovascular coupling, vascular reactivity, oxygenation and blood propagation. Temporal relationship between BOLD fluctuations in different parts of the brain provides also a mean to investigate the blood delay information. We used the induced desaturation as a label to profile transit times through different brain areas, reflecting oxygen utilization of tissue. In this study, we aimed to compare blood flow propagation delay times between these patients and healthy subjects in areas vascularized by anterior, middle and posterior cerebral arteries. In a group comparison analysis with control subjects, BOLD changes in these areas were found to be almost simultaneous and shorter in the SCD patients, because of their increased brain blood flow. Secondly, the analysis of a patient with a stenosis on the anterior cerebral artery indicated that signal of the area vascularized by this artery lagged the MCA signal. These findings suggest that sickle cell disease causes blood propagation modifications, and that these changes could be used as a biomarker of vascular damage.

Color retinal images are used manually or automatically for diagnosis and monitoring progression of a retinal diseases. Color retinal images have large luminosity and contrast variability within and across images due to the large natural variations in retinal pigmentation and complex imaging setups. The quality of retinal images may affect the performance of automatic screening tools therefore different normalization methods are developed to uniform data before applying any further analysis or processing. In this paper we propose a new reliable method to remove non-uniform illumination in retinal images and improve their contrast based on contrast of the reference image. The non-uniform illumination is removed by normalizing luminance image using local mean and standard deviation. Then the contrast is enhanced by shifting histograms of uniform illuminated retinal image toward histograms of the reference image to have similar histogram peaks. This process improve the contrast without changing inter correlation of pixels in different color channels. In compliance with the way humans perceive color, the uniform color space of LUV is used for normalization. The proposed method is widely tested on large dataset of retinal images with present of different pathologies such as Exudate, Lesion, Hemorrhages and Cotton-Wool and in different illumination conditions and imaging setups. Results shows that proposed method successfully equalize illumination and enhances contrast of retinal images without adding any extra artifacts.

Low-dose CT (LDCT) scanning is a potential way to reduce the radiation exposure of X-ray in the population. It is necessary to improve the quality of low-dose CT images. In this paper, we propose an effective algorithm for quantum noise removal in LDCT images using shearlet transform. Because the quantum noise can be simulated by Poisson process, we first transform the quantum noise by using anscombe variance stabilizing transform (VST), producing an approximately Gaussian noise with unitary variance. Second, the non-noise shearlet coefficients are obtained by adaptive hard-threshold processing in shearlet domain. Third, we reconstruct the de-noised image using the inverse shearlet transform. Finally, an anscombe inverse transform is applied to the de-noised image, which can produce the improved image. The main contribution is to combine the anscombe VST with the shearlet transform. By this way, edge coefficients and noise coefficients can be separated from high frequency sub-bands effectively. A number of experiments are performed over some LDCT images by using the proposed method. Both quantitative and visual results show that the proposed method can effectively reduce the quantum noise while enhancing the subtle details. It has certain value in clinical application.

Ultrasound Computer Tomography (USCT), developed at KIT, is a promising new imaging system for breast cancer diagnosis, and was successfully tested in a pilot study. The 3D USCT II prototype consists of several hundreds of ultrasound (US) transducers on a semi-ellipsoidal aperture. Spherical waves are sequentially emitted by individual transducers and received in parallel by many transducers. Reflectivity volumes are reconstructed by synthetic aperture focusing (SAFT). However, straight forward SAFT imaging leads to blurred images due to system imperfections. We present an extension of a previously proposed approach to enhance the images. This approach includes additional a priori information and system characteristics. Now spatial phase compensation was included. The approach was evaluated with a simulation and clinical data sets. An increase in the image quality was observed and quantitatively measured by SNR and other metrics.

Noise removal is a crucial step to enhance the quality of ultrasound images. However, some existing despeckling methods cannot ensure satisfactory restoration performance. In this paper, an adaptive non-local means (ANLM) filter is proposed for speckle noise reduction in ultrasound images. The distinctive property of the proposed method lies in that the decay parameter will not take the fixed value for the whole image but adapt itself to the variation of the local features in the ultrasound images. In the proposed method, the pre-filtered image will be obtained using the traditional NLM method. Based on the pre-filtered result, the local gradient will be computed and it will be utilized to determine the decay parameter adaptively for each image pixel. The final restored image will be produced by the ANLM method using the obtained decay parameters. Simulations on the synthetic image show that the proposed method can deliver sufficient speckle reduction while preserving image details very well and it outperforms the state-of-the-art despeckling filters in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Experiments on the clinical ultrasound image further demonstrate the practicality and advantage of the proposed method over the compared filtering methods.

Many patch-based techniques in imaging, e.g., Non-local means denoising, require tuning parameters to yield optimal results. In real-world applications, e.g., denoising of MR images, ground truth is not generally available and the process of choosing an appropriate set of parameters is a challenge. Recently, Zhu et al. proposed a method to define an image quality measure, called Q, that does not require ground truth. In this manuscript, we evaluate the effect of various parameters of the NL-means denoising on this quality metric Q. Our experiments are based on the late-gadolinium enhancement (LGE) cardiac MR images that are inherently noisy. Our described exhaustive evaluation approach can be used in tuning parameters of patch-based schemes. Even in the case that an estimation of optimal parameters is provided using another existing approach, our described method can be used as a secondary validation step. Our preliminary results suggest that denoising parameters should be case-specific rather than generic.

Myocardial perfusion imaging using CT (MPI-CT) and coronary CTA have the potential to make CT an ideal noninvasive gate-keeper for invasive coronary angiography. However, beam hardening artifacts (BHA) prevent accurate blood flow calculation in MPI-CT. BH Correction (BHC) methods require either energy-sensitive CT, not widely available, or typically a calibration-based method. We developed a calibration-free, automatic BHC (ABHC) method suitable for MPI-CT. The algorithm works with any BHC method and iteratively determines model parameters using proposed BHA-specific cost function. In this work, we use the polynomial BHC extended to three materials. The image is segmented into soft tissue, bone, and iodine images, based on mean HU and temporal enhancement. Forward projections of bone and iodine images are obtained, and in each iteration polynomial correction is applied. Corrections are then back projected and combined to obtain the current iteration’s BHC image. This process is iterated until cost is minimized. We evaluate the algorithm on simulated and physical phantom images and on preclinical MPI-CT data. The scans were obtained on a prototype spectral detector CT (SDCT) scanner (Philips Healthcare). Mono-energetic reconstructed images were used as the reference. In the simulated phantom, BH streak artifacts were reduced from 12±2HU to 1±1HU and cupping was reduced by 81%. Similarly, in physical phantom, BH streak artifacts were reduced from 48±6HU to 1±5HU and cupping was reduced by 86%. In preclinical MPI-CT images, BHA was reduced from 28±6 HU to less than 4±4HU at peak enhancement. Results suggest that the algorithm can be used to reduce BHA in conventional CT and improve MPI-CT accuracy.

Two-dimensional color Doppler echocardiography is widely used for assessing blood flow inside the heart and blood vessels. Currently, frame acquisition time for this method varies from tens to hundreds of milliseconds, depending on Doppler sector parameters. This leads to low frame rates of resulting video sequences equal to tens of Hz, which is insufficient for some diagnostic purposes, especially in pediatrics. In this paper, we present a new approach for reconstruction of 2D color Doppler cardiac images, which results in the frame rate being increased to hundreds of Hz. This approach relies on a modified method of frame reordering originally applied to real-time 3D echocardiography. There are no previous publications describing application of this method to 2D Color Doppler data. The approach has been tested on several in-vivo cardiac 2D color Doppler datasets with approximate duration of 30 sec and native frame rate of 15 Hz. The resulting image sequences had equivalent frame rates to 500Hz.

Magnetic Resonance Imaging (MRI) offers noninvasive high resolution, high contrast cross-sectional anatomic images through the body. The data of the conventional MRI is collected in spatial frequency (Fourier) domain, also known as kspace. Because there is still a great need to improve temporal resolution of MRI, Compressed Sensing (CS) in MR imaging is proposed to exploit the sparsity of MR images showing great potential to reduce the scan time significantly, however, it poses its own unique problems. This paper revisits wavelet-encoded MR imaging which replaces phase encoding in conventional MRI data acquisition with wavelet encoding by applying wavelet-shaped spatially selective radiofrequency (RF) excitation, and keeps the readout direction as frequency encoding. The practicality of wavelet encoded MRI by itself is limited due to the SNR penalties and poor time resolution compared to conventional Fourier-based MRI. To compensate for those disadvantages, this paper first introduces an undersampling scheme named significance map for sparse wavelet-encoded k-space to speed up data acquisition as well as allowing for various adaptive imaging strategies. The proposed adaptive wavelet-encoded undersampling scheme does not require prior knowledge of the subject to be scanned. Multiband (MB) parallel imaging is also incorporated with wavelet-encoded MRI by exciting multiple regions simultaneously for further reduction in scan time desirable for medical applications. The simulation and experimental results are presented showing the feasibility of the proposed approach in further reduction of the redundancy of the wavelet k-space data while maintaining relatively high quality.

The lack of noise free Optical Coherence Tomography (OCT) images makes it challenging to quantitatively evaluate performance of image processing methods such as denoising methods. The synthetic noise free OCT images are needed to evaluate performance of image processing methods. The current synthetic methods fail to generate synthetic images that represent real OCT images with present of pathologies. They cannot correctly imitate real OCT data due to a tendency to smooth the data, losing texture information and even, pathologies such as cysts are simply smoothed away by these methods. The first aim of this paper is to use mathematical models to generate a synthetic noise free image that represent real retinal OCT B-scan or volume with present of clinically important pathologies. The proposed method partitions a B-scan obtained from real OCT into three regions (vitreous, retina and choroid) by segmenting the inner limiting membrane (ILM) and retinal pigment epithelium (RPE) surfaces as well as cysts regions by medical experts. Then retina region is further divided into small blocks. Different smoothness functions are used to estimate OCT signals in vitreous, choroid and cyst regions and in blocks of retina region. Estimating signals in block resolution enables our proposed method to capture more textural information by using a simple mathematical model (smoothness function) and using annotated cyst enables our method to model cyst pathology accurately. The qualitative evaluations show that proposed method generates more realistic B-scans with present of pathologies and textural information than other methods.

Echocardiography provides valuable information to diagnose heart dysfunction. A typical exam records several minutes of real-time cardiac images. To enable complete analysis of 3D cardiac strains, 4-D (3-D+t) echocardiography is used. This results in a huge dataset and requires effective automated analysis. Ultrasound speckle tracking is an effective method for tissue motion analysis. It involves correlation of a 3D kernel (block) around a voxel with kernels in later frames. The search region is usually confined to a local neighborhood, due to biomechanical and computational constraints. For high strains and moderate frame-rates, however, this search region will remain large, leading to a considerable computational burden. Moreover, speckle decorrelation (due to high strains) leads to errors in tracking. To solve this, spatial motion coherency between adjacent voxels should be imposed, e.g., by averaging their correlation functions.¹ This requires storing correlation functions for neighboring voxels, thus increasing memory demands. In this work, we propose an efficient search using PatchMatch, ² a powerful method to find correspondences between images. Here we adopt PatchMatch for 3D volumes and radio-frequency signals. As opposed to an exact search, PatchMatch performs random sampling of the search region and propagates successive matches among neighboring voxels. We show that: 1) Inherently smooth offset propagation in PatchMatch contributes to spatial motion coherence without any additional processing or memory demand. 2) For typical scenarios, PatchMatch is at least 20 times faster than the exact search, while maintaining comparable tracking accuracy.

Respiratory motion is a challenging factor for image acquisition, image-guided procedures and perfusion quantification using contrast-enhanced ultrasound in the abdominal and thoracic region. In order to reduce the influence of respiratory motion, respiratory correction methods were investigated. In this paper we propose a novel, cluster-based respiratory correction method. In the proposed method, we assign the image frames of the corresponding respiratory phase using spectral clustering firstly. And then, we achieve the images correction automatically by finding a cluster in which points are close to each other. Unlike the traditional gating method, we don’t need to estimate the breathing cycle accurate. It is because images are similar at the corresponding respiratory phase, and they are close in high-dimensional space. The proposed method is tested on simulation image sequence and real ultrasound image sequence. The experimental results show the effectiveness of our proposed method in quantitative and qualitative.