1.1.

Burden of Ovarian Cancer

Despite concerted efforts to improve patient outcomes, ovarian cancer remains the deadliest gynecologic malignancy in the United States. While ovarian cancer is not exceedingly common, the disease maintains a high mortality rate, with median 5-year survival less than 45%.¹ One cause of this is the fact that ovarian cancer can grow to a large size before causing signs or symptoms, leading to a high proportion of advanced disease at the time of detection. In fact, a large majority of patients have already experienced spread of their disease to local or distant tissues at initial diagnosis, resulting in a significantly poorer prognosis.²

This insidious pattern of disease progression has led to strong interest in the area of ovarian cancer screening, with the ultimate goal of identifying early-stage tumors, while the patient is still asymptomatic, allowing more effective treatment. Various screening modalities have been investigated to reduce the burden of the disease including physical examination, transvaginal ultrasound, and serum tumor marker measurement (most commonly Ca-125).³ Other screening tests and multimodal protocols have also been investigated; however, at this time, no routine screening is recommended in average-risk patients.⁴ As such, there remains a strong need for a high-quality, minimally invasive modality for effective detection of early-stage ovarian malignancies.

1.2.

Optical Coherence Tomography

Optical coherence tomography (OCT) is an interferometric imaging technique first introduced in 1991⁵ that yields depth-resolved, high-resolution images of tissue, providing information about the tomography and microstructure. Historically, OCT has been successfully applied to biological imaging in the human eye,^6–8 the lung,^9,10 the esophagus,¹¹ the coronary artery,^12,13 in addition to a number of other organs, including the ovaries.^14–18 The physical principle of OCT systems is similar to that of ultrasound, except that OCT systems measure time-resolved backscattered light instead of sound waves.¹⁹ A complicating factor of OCT is the depth dependence of the system performance. Lateral resolution varies throughout the sample depth. In addition, the axial resolution degrades in deeper tissue as a result of breaking the assumption of single-scattering of light; furthermore, absorption by the tissue attenuates the signal. Ultimately, the image statistics vary as a function of depth, which can frustrate attempts at quantitative analysis. Despite these drawbacks, OCT is a widely applied and robust approach to characterizing tissue microstructure. In particular, OCT has shown great potential for disease diagnostics and tissue classification in the ovaries by imaging a wealth of microstructural features, including the stroma, epithelium, and collagen.^14–17,20

A myriad of different image analysis techniques have recently been investigated in the scope of classifying tissue health based on OCT images. Some examples include structure and texture analysis,^21–25 convolutional neural networks,^13,26,27 and other machine-learning techniques.^10,28,29 Quantitatively characterizing tissue with such approaches has shown great promise as a diagnostic aid. One important step to test and evaluate the suitability of different methods relies on mouse models, which are critical to provide a systematic control in which to observe biological variations. However, in the scope of ovarian OCT imaging, the mouse ovaries must first be separated from the image background due to the small organ size relative to a typical OCT system field of view. This process, known as image segmentation, allows the relevant image content to be extracted and analyzed, preventing corruption from background features. The need for segmentation is not unique to ovarian OCT imaging: segmentation is a common challenge in medical imaging.^30,31 Unfortunately, pre-existing solutions are tuned to a given application and do not translate well between imaging modalities.³² Segmentation can be accomplished either manually or automatically. While accurate, manual segmentation is time-intensive, particularly when OCT yields three-dimensional (3-D) data. Also of importance, appropriate preprocessing is required to suppress noise to best segment the image. Much work has investigated noise-reduction and automated segmentation for retinal OCT imaging;³³ however, little has considered the application to the ovaries, which exhibit higher structural variance and inherent inhomogeneity than the retina. Hence, to efficiently evaluate quantitative analysis methods for ovarian OCT imaging, finding an effective approach to automatic segmentation is critical. A robust segmentation algorithm would widely inform the field of OCT imaging as the application has expanded to organs, such as the esophagus,¹¹ colon,³⁴ and coronary artery,¹² where segmentation is required to quantitatively assess tissue health.^35,36 Here, we evaluate a set of preprocessing techniques and segmentation algorithms for the purpose of segmenting OCT images of the ovaries.

1.3.

Image Segmentation Methods

Many different approaches to segmentation have been proposed in the scope of medical image processing.^30,32,33 These methods can be separated into different groups, depending on the underlying mathematics involved (Fig. 1). Classical segmentation techniques partition the image into nonoverlapping, continuous, segments based on the value of some features, such as brightness or texture.^37–39 Pixel classification methods are an extension of classical approaches, albeit the constraint on region continuity is relaxed. Thresholding the image remains one of the simplest and most intuitive approaches to segmentation. In this case, a histogram is created to cluster pixels according to some feature, such as pixel intensity or a texture parameter. The thresholds are selected to partition the histogram, where the pixels residing in a given partition are assigned a label for segmentation. Depending on the feature represented in the histogram, thresholding can be used to segment based on edges or regions; other methods use a combination of the two and are referred to as hybrid algorithms.⁴⁰ For example, the watershed algorithm is a widely used technique where regions of different classes are first seeded based on finding maxima in a distance-transform of the image; the regions are then grown until the sources meet.⁴¹ The resulting boundary between sources is taken then as the segmentation.

Fig. 1

Image segmentation algorithms can be decomposed into different classes, depending on the mathematics involved. We test seven of the most commonly found techniques for medical image segmentation, each of which belongs to a different class of algorithm.

Pattern recognition algorithms can perform segmentation by identifying inherent structure in an image. These algorithms can either be supervised or unsupervised; in either case, the algorithm identifies patterns in the image, which are then used to partition the image area. Supervised approaches are first trained on a set of manually segmented images that are used as reference; examples include artificial neural networks and support vector machines (SVM).^42,43 Recently, deep learning methods have found increasingly widespread application for segmentation and analysis. The most notable is the U-Net, which was developed specifically for biological image segmentation.⁴⁴ Unsupervised methods, also known as clustering methods, do not need training data; however, properly initializing the algorithm parameters is essential for accuracy.^45–47 Another class of segmentation algorithms is referred to as global optimization methods, which are based on energy minimization techniques. One such example is graph cutting, where the image is represented as an adjacency graph.^48,49 In this graph, the vertices represent pixels of an image, and the weight between two vertices is the similarity between two given pixels. The graph is then partitioned by cutting the vertex connections to create different groups. The optimization minimizes the summation of the weights that are cut, which can be thought of as minimizing the energy in the system. Another form of energy-minimization is to fit a model to the image, which takes advantage of morphologic or structural characteristics. Examples of this include fitting deformable models and parametric curves to an image.^50,51 Recently, one such technique known as the active contour model has found much success in medical image segmentation.^52–54 An active contour model is applied by initializing a so-called snake, which is a two-dimensional (2-D) path within the image. This path can be constrained with different boundary conditions and the algorithm proceeds by fitting this path to the contours in the image. This deformable path is an energy minimizing spline influenced by image forces, defined to pull it toward object contours, which are balanced by internal forces that resist deformation. Active contours may be understood as a special case of the general technique of matching a deformable model to an image by means of energy minimization. While quite robust, the active contour approach does require knowledge of the desired contour shape beforehand to properly initialize the path.

Other classes of segmentation exist that are found less frequently in medical image analysis; for example, registration-based methods, such as ATLAS warping,^55,56 other machine learning models including active appearance modeling,^42,57,58 and a method known as locally excitatory globally inhibitory oscillator network (LEGION), which is based on a biologically plausible computational framework inspired by a biological oscillator network.^59,60 In this paper, we test the performance of seven different segmentation techniques for segmenting OCT images of the ovaries: intensity thresholding, the watershed algorithm, $k$-means clustering, graph cutting, an SVM, active contour modeling, and a deep neural network. While numerous other approaches to segmentation exist, these seven methods are a representative sample of the most widespread approaches to segmentation in medical imaging, covering a wide range of the different classes of segmentation.

2.1.

Optical Coherence Tomography System

Three-dimensional OCT imaging was completed with a swept source OCT system (OCS1050SS, Thorlabs). The system operates in noncontact mode with a central wavelength of 1040 nm and spectral bandwidth of 80nm. The axial scan rate was 16 kHz, and the power on the sample was measured as 0.36 mW. The system was set to average four axial scans. The OCT system has $11\mu \mathrm{m}$ transverse resolution and $9\text{-}\mu \mathrm{m}$ axial resolution in tissue. Imaging volume was $4\mathrm{mm}(X\text{\hspace{0.17em}lateral})\times 4\mathrm{mm}(Y\text{lateral})\times 2\mathrm{mm}(Z\text{axial})$ and $750\times 752\times 512\text{pixels}$ (voxel size of approximately $5\mu \mathrm{m}\times 5\mu \mathrm{m}\times 4\mu \mathrm{m}$). The image volume was exported as a series of 2-D en face ($X-Y$) images, or slices, and saved to disk as .tif image files.

2.2.

Mouse Model

For this experiment, we used a mouse model of ovarian cancer and imaged mice of different ages, genotypes, and reproductive statuses. Females of the transgenic mouse model (TgMISIIR-TAg) spontaneously develop bilateral epithelial ovarian cancer.^61,62 Both transgenic females and their wild type female littermates were imaged. Two reproductive status groups were obtained by dosing with 4-vinylcyclohexene diepoxide dissolved in sesame oil, which induces follicular atresia mimicking postmenopause⁶³ or vehicle (sesame oil). We imaged mice at 4 and 8 weeks of age. In total, we acquired 70 images to analyze. By examining mice at different ages, reproductive status, and genotypes, we introduce biological variability into the dataset, thus creating a challenging segmentation problem.

2.3.

Image Processing

2.3.1.

Manual segmentation protocol

We established a ground truth set of segmentation masks by manually segmenting each image using the ImageJ program.⁶⁴ To do so, a given 3-D image stack was loaded, where each en face slice in the stack represented a different depth. We located the first valid image by finding the most superficial image slice where the ovaries were visible and the image was not occluded by artifacts such as strong surface back reflections. A mask was then drawn around the ovaries using the create mask tool [Fig. 2(a)]. The result was a binary mask where the value was one within the drawn region of interest and zero elsewhere. The image was saved to disk, and the process was repeated every 10 slices until the average brightness within the ovary dropped below 20% of that recorded from the first valid image. Once this step was complete, the segmentation mask was linearly interpolated between each manually segmented slice to account for the sampling step of 10 slices [Fig. 2(b)].

Fig. 2

(a) Individual slices of the OCT image stack were manually segmented using ImageJ. (b) This was repeated throughout the image depth and interpolated to yield the final segmented volume. (c) Due to the absorbing nature of tissue, as the imaging depth increases, the signal within the tissue is expected to decrease, while the signal at the edges will remain high.

We choose to segment every 10 slices to reduce the time required for a full segmentation ($\sim 100$ to 120 slices). While this action may result in a decrease in accuracy, the shape of the organs evolves roughly linearly on a small scale; thus, the approximation by sampling every 10 slices ($\text{4\hspace{0.17em}\hspace{0.17em}}\mu \mathrm{m}$, less than two axial resolution units) is reasonable. We quantified the error introduced by interpolating the segmentation mask during manual segmentation. To do so, we manually segmented an ovary using every slice in the valid region. We then compared the accuracy of this result to the segmented mask created by interpolating between every 5, 10, and 20 slices using the performance metric defined in Sec. 2.3.4.

All images were segmented by a single observer; however, there is a degree of uncertainty as the observer must decide where the segmentation boundary is. To evaluate the error inherent in manual segmentation, a different individual segmented 25 randomly selected ovaries. We compared the randomness introduced by a given observer by computing the similarity between the two results. One factor that contributes to this observer error is the decreased signal through the imaging depth. As the imaging depth increases, we expect the signal within the center of the ovaries to decrease, as incident light propagates through more volume and absorbed more. Conversely, light striking the edge of the roughly spherical ovaries will remain high, as it undergoes relatively less attenuation [Fig. 2(c)].

2.3.2.

Preprocessing techniques

Preprocessing is an essential step of OCT image analysis due to the speckle noise intrinsic to all OCT images.⁶⁵ The presence of speckle noise, among other sources of noise, reduces image quality and can prohibitively frustrate some analysis techniques, such as texture analysis^22,29,66 and boundary identification.³³ In the body of literature, nearly all OCT image analysis methods consist of a preprocessing step to suppress electronic and speckle noise. The most common approach is to apply a median filter; however, other filtering methods have been successfully applied, which may offer advantages in preserving spatial resolution or reducing processing time.³³ Of these possibilities, we examine five representative preprocessing techniques for the reduction of speckle noise in OCT images (Fig. 3). These include mean and median filtering with a $5\times 5\text{pixel}$ kernel size, as well as Gaussian filtering with a standard deviation of five pixels. Additionally, we applied nonlinear anisotropic filtering (50 iterations; gamma = 0.1) and low-pass filtering (thresholded at 50% frequency content).

Fig. 3

(a) Five preprocessing techniques were investigated to suppress speckle noise in OCT images of the ovaries. (b) We tested mean, (c) median, (d) low-pass, (e) Gaussian filters, (f) in addition to anisotropic diffusion filtering. Each image here is a single en face slice.

The kernel size was selected to attenuate the effect of speckle in the image. Speckle caused by multiple scattered light and electronic noise is effectively one resolution unit in size;¹⁹ therefore, our kernel size was chosen as two resolution units in size to eliminate the speckle. Given the system specifications, two resolution units correspond to $\sim 4.5\text{pixel}$, leading to the selection of a 5-pixel kernel size. This kernel size, along with the parameters for the anisotropic diffusion filtering and low-pass filtering is consistent with what is found in the literature for applications with similar lateral resolution.^5,7,67 To evaluate the performance of each preprocessing technique, we calculated the average segmentation accuracy (defined in Sec. 2.3.4) across a set of 10 randomly selected test images for the seven different segmentation approaches. We also conducted segmentation on the same set of 10 test images, no preprocessing. By averaging the results of the 10 images for each preprocessing approach and taking the ratio between the processed and unprocessed images for each segmentation approach, we compute the relative increase in performance. In addition, we recorded the computation time to compare the speed of each technique. We tested the statistical significance of the results using analysis of variance (ANOVA).

3.1.

Manual Segmentation Error

We first tested the error introduced by interpolating the segmentation mask during manual segmentation. We find that the two masks yield an accuracy of 97.7% for sampling every 20 slices, 98.2% for sampling every 10 slices, and 98.3% for sampling every 5 slices. We also examined how the choice of observer can introduce randomness into the result. Comparing manual segmentations between two observers for 10 randomly selected images yielded a similarity of 97.6%; thus, the interpolation process does not introduce more error than what is inherent to the variation in the observer. We also observed that there was no appreciably decrease in accuracy when observers randomly sampled slices through the depth of an image. This error varied minimally ($<2\%$) over the sample of 25 images, which indicates that the computed error is representative of the total population. Furthermore, the error between the segmentation algorithm results and the manually segmented slices compared with the interpolated slices had no observable difference. For each subsequent manual segmentation, we sampled every 10 slices.

3.2.

Preprocessing Techniques

The average increase in accuracy when using different preprocessing techniques (as compared with no preprocessing), as well as the required processing time, is shown in Fig. 4. The results indicate that Gaussian filtering and median filtering produce the highest average improvement, with the Gaussian filtering increasing relative accuracy by 32% when compared with segmenting an unfiltered image. Considering processing time, the Gaussian filtering is also most rapid, completing in an average time of $28.4\mathrm{ms}\pm 1.4\mathrm{ms}$; second is the low-pass filter with an average processing time of $73.6\mathrm{ms}\pm 1.2\mathrm{ms}$. Taking these results with the high statistical significance of the ANOVA testing ($p<0.001$), we can conclude that Gaussian filtering is most suitable for preprocessing in our segmentation problem.

Fig. 4

Average increase in segmentation accuracy (orange) and processing time (magenta) of different filtering techniques. The Gaussian filter performs best in both categories, while the median filter also exhibits high accuracy and rapid processing time.

3.3.

Segmentation Algorithms

Figure 5 illustrates the results of applying the seven segmentation techniques, showing the maximum segmentation accuracy throughout the image depth. While each algorithm with the exception of intensity thresholding yields high maximum accuracy ($>85\%$), the active contour method performs best, with a maximum accuracy of $94.8\%\pm 1.2\%$ on average ($p<0.01$). Evaluating the performance as a function of image depth (Fig. 6), we see that clustering and active contour modeling remain the most effective, while the other approaches suffer from high variations in accuracy throughout the depth. In particular, the watershed algorithm, SVM, and thresholding all significantly degrade in performance as a function of depth.

Fig. 5

Highest accuracy obtained by each algorithm for a single 2-D slice throughout the image depth. Active contour modeling performs the best, followed by $K$-means clustering and the SVM.

Fig. 6

Segmentation accuracy as a function of image depth. We see that an active contour model maintains high accuracy throughout the depth while thresholding; the watershed algorithm and the SVM degrade rapidly throughout the depth. Clustering and graph cutting perform reasonably well, but have a larger variation than active contours.

While clustering and graph cutting perform reasonably well, active contour modeling is most accurate. These results may be due to the high inhomogeneity of the image content for OCT images. Methods such as thresholding and the SVM classifier depend on pixel intensity; these vary throughout the image depth and across the area of the ovary due to nonuniform attenuation of light in the roughly spherical ovary; furthermore, much of the connective tissue has similar levels of intensity to the ovaries. Thus, intensity-based methods cannot discriminate well between these tissue types, nor do they adapt well to depth-dependent intensity. One potential solution could be to train the SVM or threshold based on local textural features, instead of intensity features. As the local region is likely to undergo similar changes throughout depth, much of the variance could be reduced; hence, characterizing the local texture would enable a more depth-independent feature. Furthermore, qualitatively inspecting the OCT images, we observe that the connective tissue exhibits different textural differences than the ovary; therefore, texture could be a more appropriate feature for segmentation than intensity.

While clustering and graph cutting consider pixel brightness as well, they also incorporate spatial information by weighting the pixels based on the location and enforcing connectivity between segmented regions. Thus, this additional information provides a constraint that makes the algorithm more resistant to the changes in image content as a function of depth. The watershed algorithm, which also includes regional information, does not enforce connectivity and therefore is also highly susceptible to the changes in depth. However, taken generally, we see that for the region-based techniques, the accuracy is more consistent throughout the image depth. This is not surprising, as the shape and contours of the organs will remain continuous throughout the depth, making these algorithms more robust to the variations in image content throughout the depth.

Using the U-Net for deep learning, we see a marked improvement over the traditional machine learning approach of the SVM. However, the results are not as strong as what has been observed for other biological segmentation applications of the U-Net. This could be due to the lack of sufficient training data. The results vary somewhat unpredictably throughout the depth; this may indicate that the features learned by the U-Net do not appear isotropically throughout the depth. With a more extensive dataset, these feature maps could be refined to produce a more accurate and consistent segmentation. Other potential solutions include using a generative model to simulate additional training data. These methods have frequently been used to supplement limited datasets, and this remains an objective for future research.

The active contour model further improves on the region-based techniques by seeding the algorithm with an initial segmentation boundary curve. With this starting point, the algorithm then deforms the boundary curve to fit the optimal shape, given the contents of the image. Considering that the shape of the ovary changes slowly throughout the image depth, propagating the segmentation boundary throughout the depth is highly effective by providing an initial condition very close to the final solution. As a result, we observe high accuracy throughout the entire image depth using this approach (Fig. 7). As expected, the signal within the tissue drops as the depth increases; however, the signal on the edges remains high, preserving the boundary. This may be the cause of the characteristic dip in accuracy observed in Fig. 6 for several of the segmentation approaches; the contrast at the organ boundary may decrease in the center of the imaging depth.

Fig. 7

Propagating active contour snake throughout the tissue depth maintains high accuracy for delineating the segmentation boundary throughout the tissue depth, from (a) the most superficial slice to (b) shallow, (c) midrange, and (d) deep slices. The overall signal is attenuated as the depth increases, which introduces challenges with other segmentation techniques.

Observing the effectiveness of the active contour modeling, we then considered if the segmentation could be propagated throughout the image depth by sampling every $N$ slices, instead of using every slice. In this case, the sampled result could be interpolated using the same approach as with the manual segmentation. The processing time of the approach scales linearly with the number of slices; therefore, by increasing the sampling increment, we reduce the number of slices and hence the processing time. Figure 8 illustrates the effects of increasing the sampling increment. While the processing time decreases linearly, we also observe a roughly linear decrease in the average accuracy [Fig. 8(a)] for each segmented slice; furthermore, we must consider the additional error introduced through the interpolation between slices [Fig. 8(b)], which evolves linearly as well. Considering these factors, the processing time must be balanced with the desired accuracy to determine the optimal sampling depth. For example, if 88% average accuracy throughout the image depth is sufficient, a sampling depth of eight slices would be optimal.

Fig. 8

Increasing the sampling depth for propagating the active contour through (a) the image leads to a linear decrease in accuracy, (b) as well as an added interpolation error. The processing time for an image stack decreases with an increased sampling depth, suggesting that an optimal sampling depth can yield rapid processing time while maintaining high accuracy.

Summarizing the results in Table 1, we report that active contour modeling is the most accurate approach to segmenting OCT images of ovarian tissue. The results are encouraging, showing that accurate segmentation can be achieved throughout the image depth with minimal user interaction. In addition, considering that a manual segmentation required $\sim 30\mathrm{s}$ for each slice, the time required to segment the full image stack improves by approximately two orders of magnitude. Nevertheless, we identify several objectives for future investigation. First, while we find that active contours are most effective of the tested algorithms, other segmentation approaches exist, particularly in the realm of machine learning, such as neural networks and⁶⁸ active appearance modeling.³¹ Machine learning continues to evolve and these algorithms could lead to more time-efficient and accurate segmentation. In addition, the feature vector used to train the SVM contains a relatively naive measure of the local image content by simply inspecting the local mean, median, and Gaussian average. Higher accuracies could be achieved by developing a more descriptive feature vector, such as one that incorporates texture analysis.⁶⁹ Finally, with the 3-D of OCT data, traditional segmentation approaches are slow to process, as they consider each image in the stack iteratively. Extending these algorithms to apply to a 3-D image could both increase the accuracy while improving the processing time. While processing high-dimensional data is memory-intensive, recent advances in computing have lowered this barrier, making 3-D processing more feasible and an exciting frontier in image processing.

Table 1

Summary of performance for each segmentation algorithm. Active contours provide both the highest maximum accuracy, as well as the most consistent accuracy throughout the depth of the tissue, as well as a rapid processing time. Note that the processing time for the SVM does not include training.