Machine learning techniques for ophthalmologic applications

Abstract

Major advancements in computational resources and greater research focus have allowed Deep Learning to become increasingly popular and relevant in a myriad of fields in modern age. Ophthalmology is one such field that has the potential to benefit greatly from Deep Learning, and it will be the focus for this project. This project seeks to tackle 2 ophthalmologic image classification tasks: the classification of eye fundus images according to the presence of referable diabetic retinopathy, and the classification of retinal optical coherence tomography (OCT) images according to the presence of choroidal neovascularization, diabetic macular edema and drusen. The Convolutional Neural Network (CNN) model, popularly employed for image classification problems, was investigated in this project. Experiments that looked into the effects of the type of pre-trained model used for transfer learning, data augmentation, layer freezing, and differing batch sizes were conducted. The best configurations from each experiment were applied to the final models, and the models were benchmarked against other published models. The CNN model for fundus image classification achieved an accuracy, F1 score, sensitivity, specificity, and AUC of 0.8739, 0.7362, 0.7659, 0.8741 and 0.8924 respectively on the Messidor-2 dataset, while the CNN model for OCT image classification achieved an accuracy of 0.9712 and a macro-averaged F1 score of 0.9711 on a reference test dataset. A web application prototype leveraging on the 2 CNN models to make predictions was also developed and deployed.