Development of machine learning techniques for detecting ophthalmologic conditions

Abstract

With the advancement of deep learning, transfer learning has gained traction as a method for applications to medical imaging. Ophthalmology is a field that has potential to benefit from transfer learning. This project aims to apply transfer learning on Convolutional Neural Network (CNN) models to solve 2 ophthalmologic image classification tasks: the classification of retinal images according to the presence of glaucoma, and the classification of retinal images according to the grade of diabetic retinography (none, mild, moderate, severe, proliferative). Experiments that investigated the effects of the types of source datasets used during for transfer learning were carried out. The CNN models for glaucoma detection as well as diabetic retinography detection, pretrained on ImageNet, gave the best performance and achieved an accuracy and F1 score of 0.9250 and 0.9594 respectively on the REFUGE dataset, and an accuracy and F1 score of 0.7661 and 0.5769 respectively on the Messidor-2 dataset.