This PDF file contains the front matter associated with SPIE Proceedings Volume 12468, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

MR Fingerprinting is a quantitative MR imaging technique that provides multiple tissue property maps simultaneously after a single MR scan. This scan is clinically feasible and has been robustly applied to imaging different parts of the body and various diseases. In multisite studies from multiple vendors, the MRF technique demonstrated high reproducibility for T1 and T2 relaxation time mapping despite scanner variations and other confounding factors. MRF thus offers a unique opportunity to ensure quality control of the MRI source data for the following analysis and clinical decision support by providing reliable and reproducible measurements of tissue properties. Combining the precision and sensitivity of MRF with advanced image analysis techniques may lead to a unified quantitative imaging tool for precision cancer imaging and treatment planning in multisite studies.

Lung cancer tends to be detected at an advanced stage, resulting in a high patient mortality rate. Thus, recent research has focused on early disease detection. Autofluorescence bronchoscopy (AFB) is an effective noninvasive way of detecting early manifestations of lung cancer. Unfortunately, manual inspection of AFB video is extremely tedious and error-prone, while limited effort has been expended toward potentially more robust automatic AFB lesion analysis. We propose a real-time deep-learning architecture dubbed ESFPNet for accurate segmentation and robust detection of bronchial lesions in an AFB video stream. Our approach gives the best segmentation results (mDice = 0.756, mIoU=0.624) on our AFB dataset among recent architectures. Moreover, our model shows promising potential applicability to other domains, as evidenced by its state-of-the-art (SOTA) performance on the CVC-ClinicDB,ETIS-LaribPolypDB datasets, and superior performance on the Kvasir, CVC-ColonDB datasets.

This paper proposes a deep neural network, Geographic Attention Model (GA-Net), for body composition tissue segmentation. By adding an auxiliary body area prediction task, our method exploits the rich semantic and spatial features contained in the body area and incorporates the features of both area and body composition tissue. In this way, GA-Net achieves superior performance for body composition tissue segmentation, especially for the indistinguishable boundaries of multiple tissues. And the enhanced representation ability of GA-Net also allows GA-Net to obtain well generalization performance on the limited dataset.

Epicardial (EAT) and paracardial (PAT) adipose tissues (inside and outside the pericardial sac, respectively) are thought to be associated with major adverse cardiovascular events (MACE). Our long-term goal is to include PAT and EAT in a comprehensive survival analysis of MACE. Here we developed an automated method for segmenting PAT in computed tomography calcium score (CTCS) scans. Analysts identified the top and bottom heart slices by anatomical evidence, and segmented PAT in a slice-by-slice basis. Our proposed PAT segmentation approach (DeepPAT) used preprocessing steps and a multi-class automated semantic segmentation (DeepLab-v3plus) network. Preprocessing steps incorporated filtering to reduce noise, window-leveling to draw attention to sac, and morphological operations to close gaps within mask volumes. DeepPAT was trained/tested on (30/22) CTCS scans from the University Hospitals of Cleveland. The output mask voxels were classified as either enclosed sac, PAT, or background. PAT region is further thresholded with standard fat HU range [-190, -30]. The DeepPAT showed excellent segmentation compared to ground truth (manual) with an average Dice score (82.5%±3.93) and correlation of (R=99.23%, P<<0.001). PAT volume difference was (4.08%±7.78) while the PAT mean HU value changed (2.65%±4.72). The EAT and PAT volumes had a noticeable correlation R=82.9% (P<<0.001). Volumes for MACE/no-MACE (5/17 patients) subgroups showed significance for PAT (P= 0.023), while EAT had better significance (P=0.004). Mean HU values showed less significance in both PAT (p=0.81) and EAT (p=0.18). Our research results offer valuable insights that can be utilized for cardiovascular risk assessment studies.

A variety of deep learning approaches have been proposed to automatically classify Alzheimer’s disease (AD) from medical images. However, common approaches such as traditional convolutional neural networks (CNNs), lack interpretability and are prone to overfitting when trained on small datasets. As an alternative, significantly less work has explored applying deep learning approaches to region-based features that are commonly attained from atlas partitions of known regions of interest (ROI). In this work, we combine CNNs with graph neural networks (GNNs) to jointly learn an adjacency matrix of connectivity’s between ROIs as a prior for learning meaningful features for AD prediction. We apply our method to the ADNI dataset and systematically inspect the different intermediate layers of our network using t-SNE projections that show strong separation on out-of-sample data. Finally, we show that the edge probabilities alone are sufficient to reach high classification accuracy by training a secondary random forest classifier on the adjacency matrices outputted from our network and illustrate the interpretability properties of the graphs by visualizing the feature importance for all edges.

Evaluation of bone fracture risk is important for the diagnosis and treatment of osteoporosis. Bone stiffness is a major factor in determining overall bone strength and fracture risk. With recent improvements in the spatial resolution of CT systems, it is possible to visualize bone microstructure and extract texture features. It is hypothesized that bone texture can be used to improve the assessment of bone strength compared to using bone mineral density (BMD) alone. In this work, we develop image analysis models for bone stiffness estimation utilizing deep learning (DL) features, radiomics features, and gradient structure tensors (GSTs) to estimate trabecular bone stiffness for high-resolution CT. We base our analysis on a dataset containing micro-CT images of 70 individual lumbar vertebrae. Ten trabecular bone ROIs were extracted from each vertebral body and their bone structure was segmented. The mechanical stiffness of each ROI was estimated using micro-finite element (μFE) analysis. Blur and correlated noise derived from clinical high-resolution CT systems were then added to the trabecular bone ROIs to generate simulated high-resolution CT images. A 3D residual network (ResNet) was trained to extract DL features to predict μFE-derived bone stiffness from the simulated CT images. Radiomics and GST features of bone ROIs were also computed for the same task. The prediction results for DL, radiomics, and GST features combined showed the best performance with a root mean square error (RMSE) of 2.646 N/μm and an R2 of 0.881. The performance of DL features alone was better than using BMD alone or using radiomic features alone. Additionally, incorporating orientation information from GST into the models resulted in improved accuracy. We demonstrate that μFEestimated mechanical properties of lumbar vertebral trabecular bone can be inferred from high-resolution CT images and that a combination of DL, radiomic, and GST features provides the highest prediction performance.

Bicuspid aortic valve (BAV) is a hereditary disorder that develops in the fetus at the early stages of pregnancy. Though the patient may have BAV defect at the time of birth, it may not be diagnosed until the patient becomes often symptomatic in adulthood. BAV patients are at a higher risk of aneurysm growth with a high mortality rate. Hence, measurements acquired from automated aortic segmentation would aid in faster analysis of hemodynamic parameters for better risk-stratify in BAV patients. In this work, we propose a fully automated segmentation tool using a deep learning technique for fast and accurate aortic segmentation. The 3D aorta volume was segmented based on the proposed model (U-Net++) and compared with two-dimensional (2D) deep convolutional neural network (DCNN) models (U-Net and Attention U-Net). Performance metrics such as Dice similarity coefficient (DSC), Hausdorff distance (HD), and absolute volume difference (AVD) were used for model evaluation. The proposed model reported the highest DSC of 0.88±0.02 on the dataset comprising of 114 subjects (n=91 BAV and n=23 healthy cases). The HD shows a difference in mean of 3.8mm between the manual and the predicted results. Though a limited dataset was deployed in this work, the model reports a high DSC based on 3D phase contrast (PC) magnetic resonance angiogram (MRA) (PCMRA) images obtained at a clinical setting. This fully automated approach minimizes the burdensome data analysis, data annotation cost and would aid for early diagnosis and to start individualized treatment to enhance the patient outcome.

Photoacoustic imaging is an emerging preclinical modality that uses a combination of optics and acoustics mechanisms to visualize differences in optical absorption in target imaging objects. Photoacoustic imaging is potentially suitable for visualizing vasculature, as the hemoglobin in red blood cells is a prominent heat absorber and therefore serves as a great biomarker. Due to acoustic reflection, diffraction and scattering, photoacoustic imaging is subject to artifacts when the target soft tissue is close to bone. We construct an ex vivo phantom featuring vascularized soft tissue near long bone, to facilitate evaluation of photoacoustic images and to enable future research on artifact removal in photoacoustic imaging.

Non-invasive accurate prostate cancer risk assessment is crucial in radiation treatment planning that impacts patients' quality of life. In this study, we aim to develop a radiomics model using ensemble learning with multi-parametric magnetic resonance imaging (mpMRI) to classify low-grade vs high-grade prostate lesions. We identified 112 prostate patients with biopsy findings and sampled 70% and 30% of the data as training and testing datasets. There were 1198 Radiomics features extracted from mpMRI. A combination of filter-based, wrapper-based and embedded methods was used for feature selection. Ensemble classifiers included multiple machine learning models, such as random forest, k-nearest neighbor and support vector machine, for each MRI modality. A soft voting ensemble classifier was used to achieve the final performance in the test set with 82% accuracy and 0.88 AUC.

Our work is a first step towards clinically validating 3D T1ρ MRI to be used for routine clinical diagnosis of Osteoarthritis (OA) and for measuring the progression of the disease and its response to therapy in OA clinical trials. Reproducibility of T1ρ values(ms) were first measured using a phantom and then on heathy participants using both Siemens 3T Prisma and GE 3T Architect clinical scanners. Intra- and inter-vendor variability was assessed on each patient and scanner with 3 test-retest scans.

In this work, we propose MLP-Vnet, a token-based U-shaped multilayer linear perceptron-mixer (MLP-Mixer) network, incorporating a convolutional neural network for multi-structure segmentation on cardiac magnetic resonance imaging (MRI). The proposed MLP-Vnet is composed of an encoder and decoder. Taking an MRI scan as input, the semantic features are extracted by the encoder with one early convolutional block followed by four consecutive MLP-Mixer blocks. Then, the extracted features are passed to the decoder with mirrored architecture of the encoder to form a N-classes segmentation map. We evaluated our proposed network on the Automated Cardiac Diagnosis Challenge (ACDC) dataset. The performance of the network was assessed in terms of the volume- and surface-based similarities between the predicted contours and the manually delineated ground-truth contours, and computational efficiency. The volume-based similarities were measured by the Dice score coefficient (DSC), sensitivity, and precision. The surface-based similarities were measured by Hausdorff distance (HD), mean surface distance (MSD), and residual mean square distance (RMSD). The performance of the MLP-Vnet was compared with four state-of-the-art networks. The proposed network demonstrated statistically superior DSC and superior sensitivity or precision on all the three structures to the competing networks (p-value < 0.05): average DSC of 0.904, sensitivity of 0.908 and precision of 0.902 among all structures. The best surfaceased similarities were also demonstrated by the MLP-Vnet: average HD = 3.266 mm, MSD = 0.684 mm, and RMSD = 1.487 mm. Compared to the competing networks, the MLP-Vnet showed the shortest training time (7.32 hours) inference time per patient (3.12 seconds). The proposed MLP-Vnet is capable of using reasonable number of trainable parameters to solve the segmentation task on the cardiac MRI scans more quickly and accurately than the state-ofthe- art networks. This novel network could be a promising tool for accurate and efficient cardiac MRI segmentation to assist cardiac diagnosis and treatment decision making.

We produce 2 kinds of images of spinal cord in the immediate aftermath of a localized contusive injury, noninvasively and in vivo in a rat model using remitted light from a single spatially scanned 830 nm laser. Two different processing algorithms/modalities were employed 1) the PVOH algorithm producing images based on the turbidity of the spinal cord tissue which includes the cerebrospinal fluid (CSF) and 2) Raman spectroscopy which allows images based on pH. Whereas PVOH requires relatively little data processing e.g., no separation of Raman and fluorescence emission for implementation, quantitative use of spontaneous Raman spectra to calculate e.g., pH requires careful separation and accounting for the presence of underlying fluorescence. We utilize an unbiased procedure i.e., no assumptions are made, that an unsupervised machine can execute based on the inherent differences in line widths for pure vibrational and vibronic transitions in fluid media, and the overall effect of heterogenous spectral broadening at physiological temperature. The fundamentally stronger signal to noise ratio for PVOH imaging allows much greater spatial resolution for the same collection times needed to construct the Raman based pH images. This proof of principle study is consistent with a hypothesis that fast localized change in the CSF pH, induced by the mechanical disruption of tissues during the injury, disrupts the delicate balance of chemical factors stabilizing the CSF phase stability. We suggest that this combination of new methodologies for real-time imaging of chemical and physical changes in spinal cords and other tissues could have many uses.

In this paper, we propose a scheme that includes automated extraction of thrombus regions and quantitative analysis of thrombosis in confocal laser scanning microscope (CLSM) blood flow image sequence. Making thrombosis model in animal models play an important role in the development of antithrombotic drugs and ascertaining thrombosis mechanisms. Making thrombosis model in cerebral cortex of mice is usually observed using a CLSM in the fluorescence mode. However, some small changes of thrombus regions are not easily observed in CLSM blood flow image sequences. In addition, it is not easy for researchers to quantitatively analyze the degree of thrombosis. Therefore, we propose a scheme to achieve automatic thrombosis region extraction and quantitative analysis. In which, our thrombosis region extraction method uses analysis of changing pattern of thrombosis regions in CLSM blood flow image sequence. Experimental results showed that our scheme can help biological researchers observe and analyze the changes of thrombosis in animal models and reduced the use of fluorescent thrombus markers.

In this study, we re-evaluated the attainable coincidence time resolution (CTR) performance for 3×3×3 mm³ LYSO crystals coupled to matched 3×3 mm² SiPMs. This work was motivated by potential increased sensitivity in brain positron emission tomography (PET) detector blocks that would be enabled with ultrashort CTR (<100 ps). The recent progress in silicon photomultiplier (SiPM) technology, high-frequency read-out circuits, and optimized data processing is expected to lead directly to improved performance. The 3×3×3 mm³ LYSO crystals, with all sides polished to optical quality, were optically coupled to SiPMs designed and fabricated by Fondazione Bruno Kessler (FBK). An improved high frequency read-out circuit was designed and fabricated. CTR was measured using a 22Na positron source (<10 μCi) sandwiched between two identical LYSO/SiPM/read-out circuit stacks. Our studies show that a CTR of less than 80 ps, which, to the best of our knowledge, is the shortest reported CTR for 3×3×3 mm³ LYSO crystals. The results demonstrate, for the first time, that CTR performance in 3×3×3 mm³ LYSO crystals coupled to a 3mm×3mm² SiPMs is comparable to CTRs achieved for ultra-small LYSO crystals (2×2×3 mm³) coupled to large 4 × 4 mm² SiPMs. These results prove that an array of 3×3×3 mm³ LYSO/SiPM can be used to build a next generation high performance detector block with very high packing fraction, enabling ultimately very high gamma ray detection efficiency and very high system sensitivity.

Aging of the thoracic musculoskeletal system can result in adverse impacts on lung function. Measurement of rib morphology in chest CT scans and assessment of their changes between full inspiratory, or total lung capacity (TLC), and full expiratory, or residual volume (RV), help examine the impacts of rib cage-related changes on lung function. We present new and automated methods using deep learning, multi-parametric thresholding, and topological analysis to segment and label individual ribs in chest CT scans, compute static morphological features at individual rib locations, and assess their lung volume related changes (Δ_LV) between TLC and RV scans. The method was applied on TLC and RV scans from the COPDGene Iowa cohort at baseline visits, and accuracy of rib segmentation and computed metrics were examined by comparing with manually outlined results on TLC and RV scans (n=2×20). An average Dice score of 0.93 was observed in all TLC and RV rib segmentations, and root-mean-square errors for different static and Δ_LV metrics were found between 0.7 and 4.9%. Application on a larger population (n=200) revealed a five-year loss of 6.2% (p<.001) in the trendline for Δ_LV in the anterior-posterior diameter of the 5^th rib with losses of 8.4 and 4.0% for males and females, respectively. Automation of CT-based static and Δ_LV metrics of rib morphology and significant evidence of age-related changes and sex-bias establish a novel and effective tool to investigate the influence of different risk factors and comorbidities in patients with chronic lung disease and their impacts on disease progression and clinical outcomes.

Oxygenation concentration of tissue is an important factor in culturing stem cells and in studying the therapy response of cancer cells. The hypoxia bone marrow is the site to harbor cancer cells. Thus, direct high-resolution measurements of molecular O₂ would provide powerful means of monitoring cultured stem cells and therapied cancer cells. We proposed an imaging approach to measure oxygenation concentration in deep tissues, based on the XLCT, with combined strengths of high chemical sensitivity and high spatial resolution. We have developed different biosensing films for oxygenation measurements and tested these films with X-ray luminescent experiments. We have also performed phantom experiments with multiple targets to validate the XLCT imaging system with measurements at two channels.

In this paper, we propose a novel pipeline for conducting disease quantification in positron emission tomography/computed tomography (PET/CT) images on anatomically pre-defined objects. The pipeline is composed of standardized uptake value (SUV) standardization, object segmentation, and disease quantification (DQ). DQ is conducted on non-linearly standardized PET images and masks of target objects derived from CT images. Total lesion burden (TLB) is quantified by estimating normal metabolic activity (TMAn) in the object and subtracting this entity from total metabolic activity (TMA) of the object, thereby measuring the overall disease quantity of the region of interest without the necessity of explicitly segmenting individual lesions. TMAn is calculated with object-specific SUV distribution models. In the modeling stage, SUV models are constructed from a set of PET/CT images obtained from normal subjects with manually delineated masks of target objects. Two ways of SUV modeling are explored, where the mean of mean values of the modeling samples is utilized as a consistent normality value in the hard strategy, and the likelihood representing normal tissue is determined from the SUV distribution (histogram) for each SUV value in the fuzzy strategy. The evaluation experiments are conducted on a separate clinical dataset of normal subjects and a phantom dataset with lesions. The ratio of absolute TLB to TMA is taken as the metric, alleviating the individual difference of volume sizes and uptake levels. The results show that the ratios in normal objects are close to 0 and the ratios for lesions approach 1, demonstrating that contributions on TLB are minimal from the normal tissue and mainly from the lesion tissue.

Purpose: We investigate whether femur morphology is affected by body mass (BM). To establish deformations associated with obesity, we propose an atlas estimation framework based on diffeomorphic shape mapping that relaxes the point correspondence requirement common to many conventional shape modeling approaches. Methods: The study sample consisted of femora from 18 normal weight (BMI between 20-25) and 18 obese (BMI > 30) individuals (Texas State University Donated Skeletal Collection). Bone surface models (2,500 vertices and approximately 5,000 faces) were generated from CT scans of the specimens (512x512 matrix, 0.625x0.625x0.5 mm voxels). The surface models were input to an optimization algorithm that yielded an atlas representation of shape variability consisting of a mean bone template and diffeomorphic deformations matching the template onto each specimen. The accuracy of normal weight vs. obese classification using principal atlas deformation modes was established in leave-one-out experiments with Support Vector Machine (SVM) classifier. Results: We achieved 75% classification accuracy in leave-one-out SVM experiments, indicating the possibility of functional skeletal adaptations to increased body mass. By visualizing the bone surface deformation given by the SVM classification direction, we found that morphological alterations associated with obesity might include relative thickening of the femoral neck and the trochanters, and retroversion of the femoral head. Conclusions: The landmark-free atlas estimation algorithm enabled detection of morphological femur variants that might be predictive of elevated body mass.

Narrow-band imaging (NBI) bronchoscopy offers enhanced visualization of microvascular structures in the lung’s epithelium (airway walls). Recent studies suggest that such vessels are helpful in predicting the invasiveness of bronchial lesions. In particular, Shibuya characterized pathological features of lesions and studied their relationship with specific histological stages of lung cancer. We propose a method for identifying these vascular patterns using a small expert-labeled dataset. Our approach is based on a few-shot learning method using a Siamese network to learn and distinguish pathological features of the bronchial vasculature. We achieved better intra-class clustering and inter-class separation in the embedding space compared to a baseline CNN classifier. Further, a 25% increase in the overall accuracy was obtained during testing.

1000 fps HSA enables visualization of flow details, which may be important in accurately guiding interventional procedures; however, single-plane imaging may lack clear visualization of vessel geometry and flow detail. The previously presented high-speed orthogonal biplane imaging may overcome these limitations but may still result in foreshortening of vessel morphology. In certain morphologies, acquiring two non-orthogonal biplane projections at multiple angles can provide better flow detail rather than a standard orthogonal biplane acquisition. Flow studies of aneurysm models were performed, where simultaneous biplane acquisitions at various angles separating the two detector views allowed for better evaluation of morphology and flow. 3D-printed, patient-specific internal carotid artery aneurysm models were imaged with various non-orthogonal angles between the two high-speed photon-counting detectors (7.5 cm x 5 cm FOV) to provide frame-correlated simultaneous 1000-fps image sequences. Fluid dynamics were visualized in multi-angled planes of each model using automated injections of iodine contrast media. The resulting dual simultaneous frame-correlated 1000-fps acquisitions from multiple planes of each aneurysm model provided improved visualization of complex aneurysm geometries and flow streamlines. Multi-angled biplane acquisitions with frame correlation allows for further understanding of aneurysm morphology and flow details: additionally, the ability to recover fluid dynamics at depth enables accurate analysis of 3D flow streamlines, and it is expected that multiple-planar views will enable better volumetric flow visualization and quantification. Such better visualization has the potential to improve interventional procedures.

3D hemodynamic distributions are useful for the diagnosis and treatment of aneurysms. Detailed blood-flow patterns and derived velocity maps can be obtained using 1000 fps High Speed Angiography (HSA). The novel orthogonal Simultaneous Biplane High-Speed Angiography (SB-HSA) system enables flow information to be quantified in multiple planes, and with additional components of flow at depth, accurate 3D flow distributions are available. Computational Fluid Dynamics (CFD) is the current standard for derivation of volumetric flow distributions, but obtaining solution convergence is computationally expensive and time intensive. More importantly, matching in-vivo boundary conditions is non-trivial. Therefore, an experimentally derived 3D flow distribution method could offer realistic results with less computation time. Using SB-HSA image sequences, 3D X-Ray Particle Image Velocimetry (3D-XPIV) was explored as a new method for assessing 3D flow. 3D-XPIV was demonstrated using an in-vitro setup, where a patient-specific internal carotid artery aneurysm model was attached to a flow loop, and an automated injection of iodinated microspheres was used as a flow tracer. Two 1000 fps photon-counting detectors were placed orthogonally with the aneurysm model in the FOV of both planes. Frame-synchronization of the two detectors made correlation of single-particle velocity components at a given timepoint possible. With frame-rates of 1000 fps, small particle displacements between frames resolved realistic timevarying flow, where accurate velocity distributions depended on near-instantaneous velocities. 3D-XPIV velocity distributions were compared to CFD velocity distributions, where the simulation boundary conditions matched the in-vitro setup. Results showed similar velocity distributions between CFD and 3D-XPIV.

A significant challenge regarding the treatment of aneurysms is the variability in morphology and analysis of abnormal flow. With conventional DSA, low frame rates limit the flow information available to clinicians at the time of the vascular intervention. With 1000 fps High-Speed Angiography (HSA), high frame rates enable flow details to be better resolved for endovascular interventional guidance. The purpose of this work is to demonstrate how 1000 fps biplane-HSA can be used to differentiate flow features, such as vortex formation and endoleaks, amongst patient-specific internal carotid artery aneurysm phantoms pre- and post-endovascular intervention using an in-vitro flow setup. The aneurysm phantoms were attached to a flow loop configured to a carotid waveform, with automated injections of contrast media. Simultaneous Biplane High-Speed Angiographic (SB- HSA) acquisitions were obtained at 1000 fps using two photon-counting detectors with the respective aneurysm and inflow/ outflow vasculature in the FOV. After x-rays were turned on, the detector acquisitions occurred simultaneously, during which iodine contrast was injected at a continuous rate. A pipeline stent was then deployed to divert flow from the aneurysm, and image sequences were once again acquired using the same parameters. Optical Flow, an algorithm that calculates velocity based on spatial-temporal intensity changes between pixels, was used to derive velocity distributions from HSA image sequences. Both the image sequences and velocity distributions indicate detailed changes in flow features amongst the aneurysms before and after deployment of the interventional device. SBHSA can provide detailed flow analysis, including streamline and velocity changes, which may be beneficial for interventional guidance.

During intracranial aneurysm (IA) treatment with flow diverters (FDs), the device/parent artery diameter ratio may influence the ability of the device to induce aneurysm occlusion. We propose to investigate a method for optimal FD selection using Angiographic Parametric Imaging (API) and data driven methods. We selected 379 pre-treatment angiographic sequences of IAs with known occluded/unoccluded status at six months follow-up. For each IA, we extracted six API imaging biomarkers, Time to Peak, Peak Height, Mean Transit Time, Area Under the Curve, Maximum Inflow Gradient, and the Cross Correlation between the time dilution input function and the aneurysm dome. Each IA biomarker was normalized to the inlet equivalent parameter and divided in four intervals using quartiles of the observed biomarker distribution. The ratio of the IA proximal artery to nominal device diameters was used as a parameter for device selection. Based on the nominal device/proximal artery diameter ratios(r), four FD situations were considered: significantly undersized (r ≤ 0.8), undersized (0.8 < r ≤ 1), oversized (1 < r ≤ 1.2), and significantly oversized (1.2 < r). For each parameter and device ratio combination, we recorded an estimated risk of treatment failure as a percent of unoccluded IAs to the total number of treated cases within determined subgroups. This retrospective analysis of the occlusion data demonstrated insightful trends between imaging biomarkers associated with increased IA inflow, selection of the device, and failure to achieve complete occlusion at six months, indicating excellent potential for this approach to be used for intraoperative device selection.

Purpose: Previous studies have demonstrated the efficacy of contrast dilution gradient (CDG) analysis in determining large vessel velocity distributions from 1000 fps high-speed angiography (HSA). However, the method required vessel centerline extraction, which made it applicable only to non-tortuous geometries using a highly specific contrast injection technique. This study seeks to remove the need for a priori knowledge regarding the direction of flow and modify the vessel sampling method to make the algorithm more robust to non-linear geometries. Materials and Methods: 1000 fps HSA acquisitions were obtained in vitro with a benchtop flow loop using the XCActaeon (Varex Inc.) photon-counting detector, and in silico using a passive-scalar transport model within a computational fluid dynamics (CFD) simulation. CDG analyses were obtained using gridline sampling across the vessel, and subsequent 1D velocity measurement in both the x- and y-directions. The velocity magnitudes derived from the component CDG velocity vectors were aligned with CFD results via co-registration of the resulting velocity maps and compared using mean absolute percent error (MAPE) between pixels values in each method after temporal averaging of the 1-ms velocity distributions. Results: Regions well-saturated with contrast throughout the acquisition showed agreement when compared to CFD (MAPE of 18% for the carotid bifurcation inlet and MAPE of 27% for the internal carotid aneurysm), with respective completion times of 137 seconds and 5.8 seconds. Conclusions: CDG may be used to obtain velocity distributions in and surrounding vascular pathologies provided the contrast injection is sufficient to provide a gradient, and diffusion of contrast through the system is negligible.

Tissue engineering (TE) holds promise for generating lab-grown patient specific organs which can provide: (1) effective treatment for conditions that require volumetric tissue transplantation and (2) new platforms for drug testing. Even though volumetric structural information is essential for confirming successful organ maturation, TE protocol designs are currently informed through destructive and 2D construct assessment tools (e.g. histology). X-ray phase-contrast computed-tomography (PC-CT) can generate non-destructive, high resolution, 3D density maps of organ architecture. In this work, PC-CT is used as new imaging tool for guiding two TE protocols currently at the in-vitro testing stage. The first (1) involves cell-repopulation of an oesophageal scaffold, with the aim of using the regenerated construct for treating long-gap oesophageal atresia, whilst for the second (2) a lung-derived scaffold is populated with islets for regenerating a pancreas, with the “repurposed” lung offering a platform for diabetes drug testing. By combing 3D images and quantitative information, we were able to perform comprehensive construct evaluation. Specifically, we assessed volumetrically: (1) the cell-distribution within the regenerated oesophagi and (2) islet integration with the vascular tree of the lung-derived scaffold. This new information was proven to be essential for establishing corresponding TE protocols and enabled their progression to more advanced scale-up models. We are confident that PC-CT will provide the novel insights necessary to further progress TE protocols, with the next step being in-vivo testing. Crucially, the non-destructive nature of PC-CT will allow in-vivo assessments of TE constructs following their implantation into animal hosts, to investigate their successful integration.

This study aimed to determine the potential of pericoronary adipose tissue (PCAT) radiomics in non-invasively identifying thin cap fibroatheroma (TCFA) in coronary CTA images. TCFA is a precursor of plaque rupture and can only be visualized using intravascular OCT (IVOCT). Our findings showed that PCAT radiomics can be used to identify TCFA as assessed by IVOCT, potentially improving the accuracy of risk assessment and treatment planning. This is the first study to demonstrate the use of PCAT radiomics for non-invasive identification of TCFA.

We extracted features of fat depots from computed tomography calcium score (CTCS) images to predict a future major adverse cardiovascular event (MACE). Our work builds on two observations: 1) Agatston score for coronary calcifications is known to be predictive and 2) studies have shown association of epicardial adipose tissue (EAT) volume with MACE. We extracted many features of fat depots (fat-omics) and used feature assessments in modeling to predict MACE. We used time-to-event Cox model with an elastic net regularization to identify the most compelling fat-omics features, including morphological (e.g., volume and thickness) and intensity statistics (e.g., mean and max HU). We collected and engineered EAT features from a 6-year cohort study of 339 individuals (58.7%MACE) from the University Hospitals Cleveland. The cohort was MACE-enriched to balance data and to enable precise determination of best features. We found that body mass index (BMI) was not a good surrogate for EAT volume, as the correlation was minimal. The 2-year ROC result of fat-omics model was superior to other univariate models (i.e., BMI, EAT volume, and Agatston), with AUC=0.72 compared to (0.56, 0.54, and 0.57), respectively. In addition, high- and low-risk stratification was improved. In a further experiment using 166 zero-Agatston cases with 59%MACE, fat-omics model outperformed BMI or EAT. Fat-omics had AUC=0.66 compared to (0.56,0.49), respectively. Promising results indicate the importance of EAT fat-omics over traditional BMI, EAT volume, and Agatston score. Fat-omics with calcifications analyses may significantly improve MACE prediction from CTCS images.

Small liver lesions common to colorectal liver metastases (CRLMs) are challenging for convolutional neural network (CNN) segmentation models, especially when we have a wide range of slice thicknesses in the computed tomography (CT) scans. Slice thickness of CT images may vary by clinical indication. For example, thinner slices are used for presurgical planning when fine anatomic details of small vessels are required. While keeping the effective radiation dose in patients as low as possible, various slice thicknesses are employed in CRLMs due to their limitations. However, differences in slice thickness across CTs lead to significant performance degradation in CT segmentation models based on CNNs. This paper proposes a novel unsupervised attention-based interpolation model to generate intermediate slices from consecutive triplet slices in CT scans. We integrate segmentation loss during the interpolation model's training to leverage segmentation labels in existing slices to generate middle ones. Unlike common interpolation techniques in CT volumes, our model highlights the regions of interest (liver and lesions) inside the abdominal CT scans in the interpolated slice. Moreover, our model's outputs are consistent with the original input slices while increasing the segmentation performance in two cutting-edge 3D segmentation pipelines. We tested the proposed model on the CRLM dataset to upsample subjects with thick slices and create isotropic volume for our segmentation model. The produced isotropic dataset increases the Dice score in the segmentation of lesions and outperforms other interpolation approaches in terms of interpolation metrics.

Spectral micro-CT shows great potential to provide accurate material composition by utilizing the energy dependence of x-ray attenuation in different materials. This is especially well-suited for pre-clinical imaging using nanoparticle-based contrast agents in situations where quantitative material decomposition helps probe processes which are otherwise limited by poor soft tissue contrast. Our group has developed multiple generations of pre-clinical prototype PCCT systems and applied them in cancer and cardiac studies using nanoparticle contrast agents. This work aims to describe and assess the performance of a hybrid system for ex vivo high-resolution micro-CT using photon counting and energy integrating detectors. Both phantom and ex vivo mouse micro-CT data were reconstructed using our iterative, multi-channel algorithm based on the split Bregman method and regularization with rank-sparse kernel regression. A post-reconstruction spectral decomposition method was used. The system is capable of high resolution (15.6 lp/mm, 10% MTF) tomographic imaging. Despite the anti-coincidence corrections, the spectral performance of the PCD is, however, not perfect. Preliminary results show that adding energy integrating data to the PCD scan reduces the prevalence of certain PCD-specific artifacts and offers the potential for various hybrid approaches to PCD corrections. We also show that our spectral hybrid micro-CT separates calcified plaques from the iodine accumulation in our atherosclerosis mouse model. This is not possible in the EID-based CT imaging. Such hybrid spectral micro-CT will benefit both nanotechnology and imaging developments by providing an ex vivo high resolution imaging method that can validate experiments in theranostics.

Ultra-high dose-rate radiotherapy (FLASH-RT) shows the potential to eliminate tumors while sparing healthy tissues. Current FLASH-RT preclinical animal studies either euthanize animals for histological assessment or use blood tests and cytokine assays to evaluate normal tissue complications. Assessing the progression of complications in situ in live animals with a non-invasive, high-resolution, and sensitive diagnostic method is desired. This study demonstrated using in vivo respiratory-gated micro-computed tomography (micro-CT) to characterize the progression of irradiation-induced pulmonary complications caused by conventional and FLASH-RT in free-breathing mice. Twelve healthy male C57BL/6 mice completed baseline micro-CT scans. Mice were equally separated into three groups that received different treatments targeting the lungs. Treatments administered included no irradiation, 10 MV x-ray FLASH-RT, and 10 MV conventional radiotherapy with a single fraction 15 Gy prescribed dose. Post-treatment, chest cavities of mice were imaged by noninvasive in vivo prospective respiratory-gated micro-CT at 2, 4, 6, 9, and 12 weeks. The image acquisition was triggered using the measured respiratory signal to produce images representing end expiration and peak inspiration. Lung volume and lung CT number were measured for both respiratory phases to evaluate functional residual capacity and tidal volume. Micro-CT images revealed that two mice developed pneumonitis post-treatment after receiving radiotherapy. Here we demonstrated an imaging method to characterize the progression of radiation-induced pulmonary side effects in free-breathing animals.

We track the impact of the current global shortage of iodinated contrast agents on radiology operations at a major US healthcare system. Using repurposed software infrastructure of a commercial AI-based image analysis vendor (Aidoc Medical, Tel Aviv, Israel), we analyzed daily volumes of radiology service request data for a total of 17,061 Computed Tomography (CT) exams before and during the contrast agent shortage (both comprising 04/01/2022 through 07/01/2022), namely 2,407 CT Pulmonary Angiography (CTPA), 3,811 non-angiography Contrast-Enhanced Thoracic CT (CE-TCT), and, for comparison, 10,843 non-contrast head CT exams. Specifically, we compared two observational periods, namely (i) a pre-shortage control period from 04/14/22 through 05/05/2022, and (ii) a contrast shortage period from 05/21/2022 through 06/11/2022. A percentage change metric of case volumes was calculated, where we report relative changes with regard to a baseline measurement period from 04/01/2022 through 04/14/2022. The two observational periods were compared for statistically significant differences. Case volumes of contrast-enhanced CT scans dropped from baseline during the contrast agent shortage period, namely by 60.66%±23.33% for CE-TCT and 42.88%±20.22% for CTPA, respectively, where statistical differences between observational periods were highly significant (p < 10−4). Our results suggest a significant reduction of contrast-enhanced chest CT exams during the observed global contrast agent shortage, where CTPA exams were slightly less affected than other non-angiography contrast-enhanced chest CT studies. We conclude that data tracking using repurposed AI image analysis service software infrastructure can quantify effects of unexpected healthcare challenges on radiology operations, such as during the observed global contrast agent shortage.

Three-dimensional (3D) transperineal ultrasound (TPUS) is a valuable imaging tool for evaluating patients with a variety of pelvic floor disorders, including pelvic organ prolapse (POP). Patients with POP have abnormal descent of one or more pelvic organs (i.e., bladder, uterus, vagina) through the levator hiatus, which is often experienced by the patient as a persistent bothersome bulge protruding from the vaginal opening. The enlargement of the hiatal opening measured in the plane of minimal hiatal dimensions (PMHD), has been used as an indication for POP severity. Manually measuring the size of the levator hiatus in 3D TPUS images can be challenging and requires expertise and training and is timeconsuming. Hence a fully automated method for estimating the dimensions of hiatal opening is highly desirable. To this end, we developed a fully automated method to segment the levator hiatus from the PMHD based on the nnU-Net model framework. We trained, validated, and tested on a total of 252 3D US images from 138 patients that may have POP as determined by the pelvic organ prolapse quantification (POP-Q) system. As a benchmark comparison, we compared the nnU-Net to a vanilla U-Net whose hyperparameters were manually tuned. Model performance was determined using Dice similarity coefficient (DSC) and levator hiatus width, length, and area by comparing the model segmentations to manual segmentations. The nnU-Net achieved a DSC of 93.1%±3.3%, absolute width difference of 2.3mm±1.7mm, absolute length difference of 2.6mm±2.5mm and absolute area difference of 1.8cm²±1.3cm².

X-ray luminescence computed tomography (XLCT) is a hybrid molecular imaging modality combining the merits of both X-ray imaging (high spatial resolution) and optical imaging (high sensitivity to tracer nanophosphors). Narrow X-ray beam based XLCT imaging has been demonstrated to have the capacity of high spatial resolution imaging at the cost of the data acquisition time. We have primarily focused on improving the performance of the narrow X-ray beam based XLCT imaging. In a previous study, we proposed a scanning strategy achieved by reducing the scanning step size for improving the spatial resolution from double the X-ray beam size to close to the X-ray beam size. For the imaging speed, we recently introduced a continuous scanning scheme to replace the selective excitation scheme and used a photon counter to replace the oscilloscope to acquire measurement data, yielding a 16 times faster scanning time compared with what used in traditional XLCT systems. In addition, we developed a deep learning based XLCT reconstruction algorithm to reduce the number of projection views in a previous work. Moreover, we previously synthesized and compared biocompatible nanophosphors of distinct X-ray luminescence spectra to make multi-color XLCT imaging possible. Here, based on the previous work, we designed and built a first-of-its-kind fast and three-dimensional XLCT imaging system with the capacity of multi-wavelength measurements. A lab-made image acquisition software has been developed to control the system. We have performed physical experiments and verified the system performance.

Breast cancer is the most common cancer, and early detection is important to improve survival rates. For diagnosis, new imaging modality are required. Photoacoustic imaging (PAI) is arguably the most exciting 3D molecular imaging technique, since it provides functional information on the hemoglobin distribution in breast that can be used to identify malignant lesions. PAI is an absorption technique where optical pulses are used to generate sound waves. It combines both the advantages of the high contrast of optical imaging and the deep penetration of acoustic imaging. We have developed an extreme sensitivity optomechanical ultrasound sensor. This sensor enables broadband, high-resolution PAI and present great future promise in advancing breast cancer screening.

First carpometacarpal osteoarthritis (CMC-1 OA) is the most common form of OA in the hand, disproportionately affecting more women than men. The risk factors predisposing women to CMC-1 OA remain poorly understood. Fourdimensional CT (4DCT) imaging coupled with four-dimensional ultrasound (4DUS) imaging was utilized in this study to assess sex differences in joint kinematics and laxity. A male to female ratio of 1:2 CMC-1 OA patients were recruited for this preliminary study. 4DCT and 4DUS images were collected while patients performed primary thumb motions without loading on the joint. Kinematic models of the CMC-1 joint were developed to assess differences in joint kinematics between men and women. Ligament recruitment patterns at the thumb joint were assessed using our 4DUS system. The developed biomechanical models presented joint motion accurately throughout all performed motions. Current work is focused on evaluating the biomechanical risk factors that predispose women to CMC-1 OA. Anticipated results include assessing sex differences in CMC-1 joint bone morphology, the degree of joint centroid translation, joint space narrowing, as well as changes in joint congruency throughout each motion. In addition, changes in length of the dorsoradial ligament throughout thumb motion will be measured from the 4D ultrasound images collected. To the best of our knowledge, this is the first study utilizing 4DCT and 4DUS in tandem to assess thumb joint kinematics and ligament recruitment patterns. This work is a step forward in understanding the biomechanical factors and ligament recruitment patterns that result in women’s increased predisposition to the development of CMC-1 OA.

It can be difficult to identify trends and perform quality control in large, high-dimensional fMRI or omics datasets. To remedy this, we develop ImageNomer, a data visualization and analysis tool that allows inspection of both subject-level and cohort-level features. The tool allows visualization of phenotype correlation with functional connectivity (FC), partial connectivity (PC), dictionary components (PCA and our own method), and genomic data (single-nucleotide polymorphisms, SNPs). In addition, it allows visualization of weights from arbitrary ML models. ImageNomer is built with a Python backend and a Vue frontend. We validate ImageNomer using the Philadelphia Neurodevelopmental Cohort (PNC) dataset, which contains multitask fMRI and SNP data of healthy adolescents. Using correlation, greedy selection, or model weights, we find that a set of 10 FC features can explain 15% of variation in age, compared to 35% for the full 34,716 feature model. The four most significant FCs are either between bilateral default mode network (DMN) regions or spatially proximal subcortical areas. Additionally, we show that whereas both FC (fMRI) and SNPs (genomic) features can account for 10-15% of intelligence variation, this predictive ability disappears when controlling for race. We find that FC features can be used to predict race with 85% accuracy, compared to 78% accuracy for sex prediction. Using ImageNomer, this work casts doubt on the possibility of finding unbiased intelligence-related features in fMRI and SNPs of healthy adolescents.

Purpose: To advance the development of radiomic models of bone quality using the recently introduced Ultra-High Resolution CT (UHR CT), we investigate inter-scan reproducibility of trabecular bone texture features to spatially-variant azimuthal and radial blurs associated with focal spot elongation and gantry rotation. Methods: The UHR CT system features 250x250 μm detector pixels and an x-ray source with a 0.4x0.5 mm focal spot. Visualization of details down to ~150 μm has been reported for this device. A cadaveric femur was imaged on UHR CT at three radial locations within the field-of-view: 0 cm (isocenter), 9 cm from the isocenter, and 18 cm from the isocenter; we expect the non-stationary blurs to worsen with increasing radial displacement. Gray level cooccurrence (GLCM) and gray level run length (GLRLM) texture features were extracted from 237 trabecular regions of interest (ROIs, 5 cm diameter) placed at corresponding locations in the femoral head in scans obtained at the different shifts. We evaluated concordance correlation coefficient (CCC) between texture features at 0 cm (reference) and at 9 cm and 18 cm. We also investigated whether the spatially-variant blurs affect K-means clustering of trabecular bone ROIs based on their texture features. Results: The average CCCs (against the 0 cm reference) for GLCM and GLRM features were ~0.7 at 9 cm. At 18 cm, the average CCCs were reduced to ~0.17 for GLCM and ~0.26 for GLRM. The non-stationary blurs are incorporated in radiomic features of cancellous bone, leading to inconsistencies in clustering of trabecular ROIs between different radial locations: an intersection-over-union overlap of corresponding (most similar) clusters between 0 cm and 9 cm shift was >70%, but dropped to <60% for the majority of corresponding clusters between 0 cm and 18 cm shift. Conclusion: Non-stationary CT system blurs reduce inter-scan reproducibility of texture features of trabecular bone in UHR CT, especially for locations <15 cm from the isocenter. Radiomic models of bone quality derived from UHR CT measurements at isocenter might need to be revised before application in peripheral body sites such as the hips.

Cerebral aneurysm (CA) rupture is one of the major causes of hemorrhagic stroke. During endovascular therapy (ET), neurointerventionalists rely on qualitative image sequences and do not have access to crucial quantitative hemodynamic information. Quantifying angiographic image sequences can provide vital information, but it is not possible to perform this in a controlled manner in vivo. Computational fluid dynamics (CFD) is a valuable tool capable of providing high fidelity quantitative data by replicating the blood flow physics within the cerebrovasculature. In this work, we use simulated angiograms (SA) to quantify the hemodynamic interaction with a clinically utilized contrast agent. SA enables extraction of time density curves (TDC) within the desired region of interest to analyze hemodynamic parameters such as time to peak (TTP) and mean transit time (MTT) within the aneurysm. We present on the quantification of several hemodynamic parameters of interest for multiple, clinically-relevant scenarios such as variable contrast injection duration and bolus volumes for 7 patient-specific CA geometries. Results indicate that utilizing these analyses provides valuable hemodynamic information relating vascular and aneurysm morphology, contrast flow conditions and injection variability. The injected contrast circulates for multiple cardiac cycles within the aneurysmal region, especially for larger aneurysms and tortuous vasculature. The SA approach enables determination of angiographic parameters for each scenario. Together, these have the potential to overcome the existing barriers in quantifying angiographic procedures in vitro or in vivo, and can provide clinically valuable hemodynamic insights for CA treatment.

Traumatic brain injuries (TBIs) are a major health risk that increases with age. Natural brain aging results in cerebral atrophy and the enlargement of the ventricular regions. The objective of this study is to investigate the effect of cerebral atrophy on brain biomechanics with subject-specific models to determine the risk of traumatic brain injury (TBI). Utilizing subjects from a longitudinal study of aging in healthy volunteers, we created subject-specific brain models of a small cohort with progressive age-related cerebral atrophy. We then simulate concussive loading conditions to study changes in brain deformation, a correlate to risk of TBI. The results display differing trends with increasing ventricle volume, with some subjects exhibiting increases and others showing decreasing strain. Additional subject simulations are needed to clarify these the causes of these trends.

Changes in brain network connectivity can be observed in schizophrenia and other psychiatric diseases. We investigate whether large-scale Extended Granger Causality (lsXGC) can capture such alterations using resting-state fMRI data. Our method utilizes dimension reduction combined with the augmentation of source timeseries in a predictive time-series model for estimating directed causal relationships among fMRI time-series. As a multivariate approach, lsXGC identifies the relationship of the underlying dynamic system in the presence of all other time-series. Here, we examine the ability of lsXGC to accurately identify schizophrenia patients from fMRI data using a subset of 31 subjects from the Centers of Biomedical Research Excellence (COBRE) data repository. We use brain connections estimated by lsXGC as features for classification. After feature extraction, we perform feature selection by Kendall’s tau rank correlation coefficient followed by classification using a support vector machine. For reference, we compare our results with cross-correlation, typically used in the literature as a standard measure of functional connectivity, and several other standard methods. Using 100 different training/test data splits with 10-fold cross-validation we obtain mean/std f1-scores of 87.40% ± 19.73% and mean Area Under the receiver operating characteristic Curve (AUC) values of 95.00% ± 13.69% across all tested numbers of features for lsXGC, which is significantly better than the results obtained with cross-correlation (AUC=54.75% ± 30.96%, f1-score=51.10% ± 27.54%), and multiple other competing methods, including partial correlation, tangent, precision, and covariance methods. Our results suggest the applicability of lsXGC as a potential imaging biomarker for schizophrenia.

The literature suggests that schizophrenia is associated with alterations in brain network connectivity. We investigate whether large-scale Augmented Granger Causality (lsAGC) can capture such alterations using restingstate fMRI data. Our method utilizes dimension reduction combined with the augmentation of source time-series in a predictive time-series model for estimating directed causal relationships among fMRI time-series. As a multivariate approach, lsAGC identifies the relationship of the underlying dynamic system in the presence of all other time-series. Here, we examine the ability of lsAGC to accurately identify schizophrenia patients from fMRI data using a subset of 31 subjects from the Centers of Biomedical Research Excellence (COBRE) data repository. We use brain connections estimated by lsAGC as features for classification. After feature extraction, we perform feature selection by Kendall’s tau rank correlation coefficient followed by classification using a support vector machine. For reference, we compare our results with cross-correlation, typically used in the literature as a standard measure of functional connectivity, and several other standard methods. Using 30 different training/test data splits with 10-fold cross-validation we obtain mean/std f1-scores of 82.89% ± 17.25% and mean Area Under the receiver operating characteristic Curve (AUC) values of 93.33% ± 12.81% across all tested numbers of features for lsAGC, which is significantly better than the results obtained with cross-correlation (AUC=78.33% ± 25.60%, f1-score=66.22% ± 24.73%), and multiple other competing methods, including partial correlation, tangent, precision, and covariance methods. Our results suggest the applicability of lsAGC as a potential imaging biomarker for schizophrenia.

Schizophrenia is associated with alternations in brain network connectivity. We investigate whether large-scale Granger Causality (lsGC) can capture such alterations using resting-state fMRI data. Our method utilizes dimension reduction combined with the augmentation of source time-series in a predictive time-series model for estimating directed causal relationships among fMRI time-series. As a multivariate approach, lsGC identifies the relationship of the underlying dynamic system in the presence of all other time-series. Here, we examine the ability of lsGC to accurately identify schizophrenia patients from fMRI data using a subset of 31 subjects from the Centers of Biomedical Research Excellence (COBRE) data repository. We use brain connections estimated by lsGC as features for classification. After feature extraction, we perform feature selection by Kendall’s tau rank correlation coefficient followed by classification using a support vector machine. For reference, we compare our results with cross-correlation, typically used in the literature as a standard measure of functional connectivity, and several other standard methods. Using 100 different training/test data splits with 10-fold cross-validation we obtain mean/std f1-scores of 84.20% ± 20.42% and mean Area Under the receiver operating characteristic Curve (AUC) values of 94.50% ± 15.24% across all tested numbers of features for lsGC, which is significantly better than the results obtained with cross-correlation (AUC=64.50% ± 33.39%, f1-score=46.67% ± 34.01%), and multiple other competing methods, including partial correlation, tangent, precision, and covariance methods. Our results suggest the applicability of lsGC as a potential imaging biomarker for schizophrenia.

Evaluation of the intra-subject reproducibility of radiomic features is pivotal but challenging because it requires multiple replicate measurements, typically lacking in the clinical setting. Radiomics analysis based on computed tomography (CT) has been increasingly used to characterize liver malignancies and liver diffusive diseases. However, radiomic features are greatly affected by scanning parameters and reconstruction kernels, among other factors. In this study, we examined the effects of diets, reconstruction kernels, and liver-to-spleen normalization on the intra-subject reproducibility of radiomic features. The final goal of this work is to create a framework that may help identify reproducible radiomics features suitable for further diagnosis and grading of fatty liver disease in nonhuman primates using radiomics analysis. As a first step, the identification of reproducible features is essential. To accomplish this aim, we retrospectively analyzed serial CT images from two groups of crab-eating macaques, fed a normal or atherogenic diet. Serial CT examinations resulted in 45 high-resolution scans. From each scan, two CT images were reconstructed using a standard B kernel and a bone-enhanced D kernel, with and without normalization relative to the spleen. Radiomic features were extracted from six regions in the liver parenchyma. Intra-subject variability showed that many features are fully reproducible regardless of liver disease status whereas others are significantly different in a limited number of tests. Features significantly different between the normal and atherogenic diet groups were also investigated. Reproducible features were listed, with normalized images having more reproducible features.

Radiation doses delivered to entire vertebral bodies are current standard practice for the growing pediatric proton craniospinal irradiation (CSI) patients who are growing children. This procedure prevents patients from developing radiation-induced growth impairment, but it will cause hematopoietic marrow suppression. We aim to develop a noninvasive method to verify radiation damage to the marrow in spine vertebrae during fractional treatment using multiple magnetic resonance imaging (MRI) scans. We identified five pediatric patients who received proton CSI treatment with prescription relative biological effectiveness doses of 36 Gy for the spine. Each patient underwent multiple MRI scans during the treatment using T1-weighted sequences. Sagittal MR images were analyzed and focused on lumbar spine regions. Multi-Gaussian models were used to fit histograms from different MR images to quantify the radiation-induced damage to the bone marrow. MR images acquired before the treatment served as the reference to ensure no radiation-induced damage was found. After the treatment started, radiation-induced fatty marrow filtration showed in the vertebral bodies. We defined the radiation-induced damage based on the ratio between fatty marrow imaging pixels and total pixels in spine marrow, L1-L5 level. Damage fractions increased rapidly when the vertebral bodies received doses between 14 Gy and 34 Gy. The maximum damage happened approximately 40 days from the treatment start. After that, bone marrow regeneration was observed, and the damage fractions decreased. The proposed method can potentially achieve adaptative proton plan modification on the fly.

Breast-conserving surgery (BCS) is an alternative approach for tumor removal in breast tumor patients. This procedure requires confirmation of the tumor margin. In this study, a high-resolution cone-beam X-ray computed tomography (CBCT) was designed and evaluated for the application of breast tumor margin delineation. The imaging system was validated with a CBCT image quality phantom for spatial resolution and contrast. In addition, a breast phantom containing multiple pathological features was investigated. Finally, a breast tumor specimen was examined on the high-resolution CBCT. The results showed that the imaging system could provide a decent spatial resolution and great contrast. The pathological features of the breast phantom were clearly visible. The margin of the excised breast specimen was well defined by the system. Overall, the high-resolution CBCT provides a promising result for the clinical application.

Brain tumors are caused by abnormal cell growth and can cause pain and reduced survival rates. The early detection of brain tumors is pivotal in improving outcomes. Recently, magnetic resonance imaging (MRI) has been widely deployed in clinics to diagnose brain lesions non-invasively and prevent patients from receiving radiation doses of diagnostic imaging modalities. Traditionally, medical oncologists and radiologists diagnose brain tumors as benign or malignant using visual analysis of MRI images. The decision-making process is labor intensive, and relies on the expertise level of physicians. Recently, deep learning has dramatically changed the landscape of oncology by enabling automatic and accurate diagnosis. While the backbones of most state-of-the-art architectures are convolutional neural networks or vision transformers, the application of graph neural networks in radiation oncology has not yet been explored. To the authors' knowledge, this is the first demonstration of using fully-automated graph-feature-based classifiers for end-to-end brain tumor detection, indicating an overall classification accuracy of 94.89%. The proposed graph-feature-based classifiers are accessible for clinical implementation and could potentially assist radiation oncologists to precisely and accurately diagnose and prognosticate brain lesions.

Finite element (FE) analysis is an exciting computational technique that permits the collection of biomechanical data. It is widely utilized in industrial engineering, anthropology, comparative anatomy, and medicine. Unfortunately, there are still many aspects of FE analysis that need to be studied in order for this technique to more effectively support biomedical research. The current study examines how material property variation influences FE data to further advance FE analysis and augment biomedical data validity. Using standardized segmentation, 3D anatomical models of whole femur structure were obtained from cadaveric CT data provided by the University at Buffalo Anatomical Gift Program. FE analysis of the model experimental groups with different elastic properties was carried out simulating physiological loading of the femur consistent with previous biomechanical experiments on the femur model system. The results revealed that minor changes in material properties of FE models yield statistically significant differences in maximum displacement, average displacement, and average strain. Regional strain disparities were especially prominent at the inferior femoral neck, medial aspect of the femoral shaft, and the distolateral femur. The results indicate that Young’s modulus variation that is smaller than the variation in Young’s modulus values between FE studies leads to significant differences in biomechanical data. Therefore, these findings underscore the necessity for careful selection of exact elastic properties that are informed by validation data when feasible and consistent for particular anatomical structures across studies in order to advance FE modeling in biomedical research.

Experimental ischemic stroke models play an important role in realizing the mechanism of cerebral ischemia and evaluating the development of pathological extent. An efficient and reliable image segmentation tool to automatically identify the infarct region in the diffusion weighted imaging (DWI) and T2-weighted MRI (T2WI) images is critical for subsequent processing applications. This paper develops an automatic infarct segmentation algorithm in both rat brain DWI and T2WI images after stroke for further evaluation of neurological damages. The proposed framework consists of four major steps including image preprocessing, image registration, image enhancement, and infarct segmentation. To achieve complete automation, the input rat brain is first divided into two hemispheres, from which the initial infarct mask is acquired after a series of image registration, image subtraction, and image enhancement processes. Subsequently, an adaptive deformable model is exploited to perform infarct region segmentation. The proposed deformable model employs two-phase level set evolution, which is regularized by a local region integration. The integration of the difference between the local intensities and the global mean intensity is restricted in the inward and outward normal directions to minimize the influence of the intensity inhomogeneity. Moreover, the time step is dynamically modified towards annealing for performance refinement. Massive MR images were utilized to evaluate this new infarct segmentation algorithm. Adequate infarct segmentation results were obtained, which outperformed other competitive methods both qualitatively and quantitatively. Our infarct segmentation framework is of potential in providing a decent tool to facilitate preclinical stroke investigation and relevant neuroscience research using DWI and T2WI images.

The aorta is the largest vessel of the human body and its pathological degenerations, such as dissections and aneurysms, can be life threatening. An automatic and fast segmentation of the aorta can therefore be a helpful tool to quickly identify an abnormal anatomy. The segmentation of the aortic vessel tree (AVT) typically requires extensive manual labor, but, in recent years, progress in deep learning techniques made the automation of this process viable. For this purpose, we tested different deep learning networks to segment the aortic vessel tree from computed tomography angiography (CTA) scans with a deep neural network consisting of an encoder-decoder architecture with skip connections and an optional self-attention block. The networks were trained on a dataset of 56 CTA scans from three different sources and resulted in Dice score similarities between 0.043−0.897. Generally, the classical U-Nets performed better than the ones containing a self-attention block, indicating that they might diminish performance for AVT segmentation. The quality of the resulting segmentations was highly dependent on the CTA image quality, especially on the contrast between the aorta and the surrounding tissues. However, the trained deep neural network can segment CTA scans well with limited computational resources and training data.

Lung cancer is the leading cause of cancer-related death among both men and women and second most commonly diagnosed cancer, accounting for 18% of the total cancer deaths world-wide [4]. Screening high-risk patients with low dose Computed Tomography (CT) can lead to earlier treatment and increase the survival rate [5]. However, cancer diagnosis remains a challenging problem due to the subtle visual differences between benign and malignant nodules in CT images. Hence, computer-aided diagnosis (CADx) systems may prove useful in assisting radiologists in the malignancy prediction task. Previously we developed a convolutional attention-based network that allows for use of pre-trained 2-D convolutional feature extractors and is extendable to multi-timepoint classification in a Siamese structure [6]. The use of a sequence of parallel 2-D CNNs in place of a 3D CNN will result in significant reduction in the number of network parameters. In this paper, we keep the same overall structure utilized in [6] including the attention mechanism. However, herein, we report on use of Efficient-Net [1] for 2-D feature extractors due to its success on the Image-Net classification challenge. Variations of the Efficient-Net B0 to B7 pretrained on Image-Net were fine-tuned and applied to NLSTx data [6] a subset of data acquired in the National Lung Screening Trial (NLST) [2]. NLSTx includes data from biopsy confirmed scans in 650 benign and 207 malignant nodules at up to 3 time points. In our study, the performance of the best Efficient- Net reached an area under ROC curve of .7896 for the benign/malignant classification.

4D Flow Magnetic Resonance Imaging (MRI) allows non-invasive assessment of cardiovascular hemodynamics through the acquisition of three-dimensional pulsatile velocities in a single scan. However, this technique is often plagued by issues of noise and low resolution. In this paper, we employed a deep learning-based super-resolution method utilizing an SR residual network (ResNet) to enhance the measurement of hemodynamic indices at a higher resolution. Our approach enables the derivation of hemodynamic parameters dependent on spatiotemporal velocity derivatives such as vorticity, circulation, and turbulent kinetic energy, which were validated using a phantom model of arterial stenosis. We also compared the deep learning approach with linear, nearest neighbor, and natural interpolation methods with a 2x upsampling factor. The results were evaluated against Computational Fluid Dynamics simulations as a reference and showed that the deep learning approach improved the accuracy of turbulent kinetic energy (TKE) and viscous energy loss at peak systole by 7% and 9%, respectively, indicating a significant enhancement over traditional interpolation methods. Additionally, herein we introduce a novel hemodynamic parameter, enstrophy, as a potential diagnostic biomarker for assessing stenosis severity. Overall, our findings suggest that deep learning is a reliable and efficient approach for predicting hemodynamic parameters from 4Dflow MRI.

Tumor-node-metastasis (TNM) classification for lung cancer is essential for appropriate treatment strategies and has been used widely in the investigation and treatment of this cancer. In TNM classification, N descriptors are one of the most important prognostic indicators and are determined by the metastatic lymph node stations. Therefore, accurate classification of lymph nodes is crucial. Thoracic contrast-enhanced Computed Tomography (CT) images represent the gold-standard modality. However, manual segmentation and classification of lymph nodes are challenges that arise from the relatively similar attenuation between lymph nodes and surrounding structures. Recent progress of convolutional neural network (CNN) has spawned research on mediastinal lymph nodes segmentation on chest CT images using CNNs. However, the previous CNN-based method did not consider the relationship between airways and lymph node locations for segmenting the thoracic N1 lymph nodes group. In this study, we investigate whether distance maps based on tracheobronchial labeling can represent the anatomy properties of the N1 lymph nodes group in volumetric CT images using the NIH open-source dataset.