Abstract
Trauma, eye diseases, or eye surgery can compromise the endothelium of the cornea and its function, which may lead to loss of transparency and, ultimately, the necessity for corneal transplantation. Non-contact in vivo imaging of the endothelial cell layer allows the assessment of its health status. This thesis presents methods based on advanced image processing, machine learning, and deep learning for the estimation of image-based biomarkers to characterize the cell architecture of the corneal endothelium.
Because of the important role that the corneal endothelium plays in human vision, it is clinically relevant to image it and quantify its health status. Thanks to specular microscopy, the endothelium can be imaged in vivo in the clinic, quickly and in a noninvase manner. The specular images could then be segmented in order to estimate several corneal parameters (endothelial cell density [ECD], hexagonality of the cells [HEX], and cell size variation [CV]), which are used to assess the state of the tissue. Because manual segmentation is very tedious and time-consuming, there is a need for automatic tools, particularly in clinical studies with a large number of subjects. However, current segmentation tools are not accurate enough and make significant mistakes, or are unable
to detect enough cells. This is due to the poor quality of the specular images, which tend to have low contrast and many noisy artifacts, particularly for pathological corneas.
Initially, a machine learning approach based on Support Vector Machines was proposed (Chapters 2 & 3), which exploited the idea of merging superpixels. Specifically, the method would start by generating an oversegmented image comprised of superpixels followed by a merging process that would evaluate all possible combinations of two and three superpixels, merging those that would form a whole cell. However, a different methodology based on Convolutional Neural Networks (CNN, Chapter 4) provided significantly better performance, thus becoming the preferred option. This methodology used a deep learning network named U-net, which was capable of inferring the cell segmentation in a fraction of a second with very high accuracy. The framework was then extended with another CNN to select the areas in the image where cells were correctly delineated (Chapter 5), which allowed for a fully-automatic approach.
Given the great performance of CNNs to solve this problem, the idea of using a single CNN to estimate the corneal parameters directly was also explored (Chapter 6). However, the performance degraded significantly, therefore consolidating the approach of using a segmentation tool to solve the problem.
In this thesis, we have presented a fully-automatic method based on deep learning that segments the specular images and estimates the corneal parameters. The ultimate goal was to apply such methodology in two clinical studies from the Rotterdam Eye Hospital. The first study dealt with the transplantation of the cornea, following the recovery of 41 patients during the first year after surgery. The second study dealt with the implantation of a drainage device in 192 glaucomatous eyes, with the purpose of studying the
effect of the device’s tube on the endothelium during the 2 years after implantation.
The fully-automatic deep learning-based method was applied to all images in both studies. In the transplantation study (Chapter 7), the proposed method was able to estimate the corneal parameters in practically all images, resulting in an estimation error that was more than three times smaller than the one from the microscope’s software, and the number of correctly segmented cells per image was three times higher in the proposed method. This was a remarkable improvement. Similarly, the method was applied to the glaucoma study (Chapter 8), achieving even better performance since these images had significantly better quality. CV and HEX, two biomarkers barely used in clinical studies due to the difficulties to automatically estimate them, were now obtained with high accuracy, and they depicted peculiar patterns during the healing process in
both studies. The clinical relevance of such patterns is still something that needs further research. Nonetheless, the work presented here opens new opportunities to study the corneal parameters in clinical studies with a high number of patients and follow-us, where manual annotations are not a feasible solution.
Because of the important role that the corneal endothelium plays in human vision, it is clinically relevant to image it and quantify its health status. Thanks to specular microscopy, the endothelium can be imaged in vivo in the clinic, quickly and in a noninvase manner. The specular images could then be segmented in order to estimate several corneal parameters (endothelial cell density [ECD], hexagonality of the cells [HEX], and cell size variation [CV]), which are used to assess the state of the tissue. Because manual segmentation is very tedious and time-consuming, there is a need for automatic tools, particularly in clinical studies with a large number of subjects. However, current segmentation tools are not accurate enough and make significant mistakes, or are unable
to detect enough cells. This is due to the poor quality of the specular images, which tend to have low contrast and many noisy artifacts, particularly for pathological corneas.
Initially, a machine learning approach based on Support Vector Machines was proposed (Chapters 2 & 3), which exploited the idea of merging superpixels. Specifically, the method would start by generating an oversegmented image comprised of superpixels followed by a merging process that would evaluate all possible combinations of two and three superpixels, merging those that would form a whole cell. However, a different methodology based on Convolutional Neural Networks (CNN, Chapter 4) provided significantly better performance, thus becoming the preferred option. This methodology used a deep learning network named U-net, which was capable of inferring the cell segmentation in a fraction of a second with very high accuracy. The framework was then extended with another CNN to select the areas in the image where cells were correctly delineated (Chapter 5), which allowed for a fully-automatic approach.
Given the great performance of CNNs to solve this problem, the idea of using a single CNN to estimate the corneal parameters directly was also explored (Chapter 6). However, the performance degraded significantly, therefore consolidating the approach of using a segmentation tool to solve the problem.
In this thesis, we have presented a fully-automatic method based on deep learning that segments the specular images and estimates the corneal parameters. The ultimate goal was to apply such methodology in two clinical studies from the Rotterdam Eye Hospital. The first study dealt with the transplantation of the cornea, following the recovery of 41 patients during the first year after surgery. The second study dealt with the implantation of a drainage device in 192 glaucomatous eyes, with the purpose of studying the
effect of the device’s tube on the endothelium during the 2 years after implantation.
The fully-automatic deep learning-based method was applied to all images in both studies. In the transplantation study (Chapter 7), the proposed method was able to estimate the corneal parameters in practically all images, resulting in an estimation error that was more than three times smaller than the one from the microscope’s software, and the number of correctly segmented cells per image was three times higher in the proposed method. This was a remarkable improvement. Similarly, the method was applied to the glaucoma study (Chapter 8), achieving even better performance since these images had significantly better quality. CV and HEX, two biomarkers barely used in clinical studies due to the difficulties to automatically estimate them, were now obtained with high accuracy, and they depicted peculiar patterns during the healing process in
both studies. The clinical relevance of such patterns is still something that needs further research. Nonetheless, the work presented here opens new opportunities to study the corneal parameters in clinical studies with a high number of patients and follow-us, where manual annotations are not a feasible solution.
Original language | English |
---|---|
Qualification | Doctor of Philosophy |
Awarding Institution | |
Supervisors/Advisors |
|
Award date | 13 Jan 2022 |
Place of Publication | Delft |
Publication status | Published - 2022 |