Motivation

More than 40% of elderly adults in India suffer from blinding eye diseases, and most government hospitals have no smart technology to detect them. This gave me the personal push to design a fully automated deep learning-based algorithm, easily deployable on PCs with limited capabilities.

Overview

Detection of eye diseases in the early stages is crucial for the prevention of blindness. Here, we develop an automated classification algorithm for the early and late stages of retinal diseases, with a focus on good recall and F1-scores for the early stage. Our work was divided into two parts: retinal OCT scan segmentation and classification. We proposed a novel moment of inertia inspired method for proper orientation and segmentation of retinal OCT images. A novel retinal flattening algorithm was proposed which utilised thresholding of OCT images, to extract the hyper-reflective RPE layer of the retina and flattened it by fitting a higher-order polynomial. This algorithm was able to flatten highly noisy OCT images as well.

After gaining experience from transfer learning models, we proposed novel deep learning-based OCT image classifier, utilising a small CNN architecture called SimpleNet. It provides better classification accuracy with 800x fewer parameters, 350x less model size and is 50x faster during testing compared to state-of-the-art deep CNNs. Unlike other papers focusing on the prediction of specific diseases, we focus on broadly classifying OCT images into needing urgent anti-VEGF therapy, needing simple routine care or normal healthy retinas.