Retinal Segmentation for Glaucoma Diagnosis Using Deep Learning

Abstract

Glaucoma is one of the leading causes of irreversible blindness around the world. One of the significant indicators of Glaucoma is the enlargement of the Optic Cup (OC) to the Optic Disc (OD). In this project, we created a computer algorithm to identify the ratio between the sizes of OD and OC to further predict the likelihood of glaucoma. The algorithm consisted of a two-step approach. The first step of the approach involved the identification and cropping of the Optic Nerve Head area, the part that contains the OD and OC in the retina. This was done using thresholding based on the brightest spot in the image. The second step was to segment the OD and OC from the cropped image. We used an image augmentation library to transform the images in our dataset to add variability. We passed the transformed images into a deep learning algorithm known as U-Net to segment the OD and OC. Our analysis in training separate deep learning models for OD and OC yielded a higher accuracy as compared to training the models for OD and OC together. Our model has the potential of performing better with greater precision to segment images for diagnosing Glaucoma. The future work includes enhancing the model by making some additional changes to increase the model’s accuracy.

Department: Computer Science 

Faculty Mentor: Dr. Dana Cobzas