AbstractGlaucoma is a high prevalence optic neuropathy that may lead to irreversible blindness. The assessment of the anterior chamber angle of the eye is recommended by international guidelines to evaluate risk factors, categorise the disease and decide treatment strategies. The clinical-standard technique, called gonioscopy, is a difficult manual examination, seldom practised in primary care settings.
The NIDEK GS-1 digital imaging device have automated several phases of gonioscopy making it available to non-expert operators and allowing to store high quality pictures of the eye region. However, images of the angle still need extensive knowledge and time to be interpreted correctly to produce a diagnosis.
We aimed to research whether deep learning systems, currently studied in many other applications in medicine and ophthalmology, could effectively support the analysis of digital gonioscopic images, enabling patient screening in primary care settings. Experienced ophthalmologists have been involved in the collection and evaluation of clinical requirements to focus our work, leading to the selection of two main image analysis tasks to investigate. The first is the semantic segmentation of anatomical layers in gonioscopic images.
We designed and developed an algorithm capable of providing, for the first time, a rich morphological information about the eye region. The algorithm can deal with specific image (e.g., peripheral blur and vignette) and ground truth (e.g., partial annotations) characteristics that would make other state-of-the-art systems ineffective, returning an overall segmentation accuracy above 90% on our test set. Moreover, the system may estimate the reliability of its results by generating uncertainty maps through Monte Carlo dropout that proved to be effective at detecting small segmentation flaws.
The second is the automatic grading of angle aperture, a measure of clinical relevance. Differently from existing literature, in which the angle aperture is estimated over large angle quadrants (90Â°-wide), we explored solutions to grade smaller angle regions (about 5Â°-wide) for an increased sensitivity at detecting local angle closures. Our results highlight potentials and limitations of this approach and provide useful hints for future development in this field. Both development phases followed inter-domain sessions with clinical and technical collaborators for the formulation of annotation protocols and the generation of ground truth. Moreover, we have studied inter-annotator variability of ground truth delineations of angle layers to provide a comprehensive context for evaluating the performance of automated systems. In particular, we could compare the performance of our segmentation model with the average per-layer variability among experts finding good correlation. Results suggest that the identification and delineation of some of the anatomical layers of the angle is difficult even to experts and that, as a consequence, the limited agreement among annotators reflects on algorithm's performance.
Despite the limited size of our annotated datasets, we proved that deep learning systems can aid the analysis of digital gonioscopic images, possibly encouraging further research in this field. Semi-automated imaging devices and automatic analysis software may offer an effective and efficient alternative to conventional gonioscopy (the current clinical-standard) and help both prevent the development of glaucoma through better screening plans, and manage its treatment.
|Date of Award
|NIDEK Technologies Srl
|Manuel Trucco (Supervisor)