Disease detection in the eye with machine learning techniques

Abstract

With AI becoming more prevalent in the healthcare industry, this project explores the integration of machine learning and deep learning techniques in the field of ophthalmology. There is a need for accurate and efficient diagnosis as the traditional approaches often require lengthy consultations and screenings.

This project attempted papilledema detection through fundus images using an anomaly detection approach with an auto-encoder network. However, the results were sub-par due to small differences between normal and diseased fundus images. The study then moved on to glaucoma detection, where a 2-part Coarse optic disc (OD) to Fine OD/OC segmentation was employed with U-Net. This approach outperformed the direct OD/OC segmentation model. Several experiments were conducted for glaucoma classification models including augmentations, addition of external datasets, using cropped optic nerve head region images and model ensemble techniques. The best performing ensemble model outperformed the 2nd place team in the REFUGE2 Challenge for glaucoma detection, showing the capability of the model.

These deep learning models were then modified with Grad-CAM to provide visual explanations on the classification decisions made. These visualisations indicated that the region outside optic disc plays an important role in the decision-making process. To tie the whole project up, the models were deployed in a web application, where users could easily upload fundus images and receive valuable insights.