Understanding inherent image features in CNN-based assessment of diabetic retinopathy

Diabetic retinopathy (DR) is a leading cause of blindness and affects millions of people throughout the world. Early detection and timely checkups are key to reduce the risk of blindness. Automated grading of DR is a cost-effective way to ensure early detection and timely checkups. Deep learning or more specifically convolutional neural network (CNN)-based methods produce state-of-the-art performance in DR detection. Whilst CNN based methods have been proposed, no comparisons have been done between the extracted image features and their clinical relevance. Here we first adopt a CNN visualization strategy to discover the inherent image features involved in the CNN's decision-making process. Then, we critically analyze those features with respect to commonly known pathologies namely microaneurysms, hemorrhages and exudates, and other ocular components. We also critically analyze different CNNs by considering what image features they pick up during learning to predict and justify their clinical relevance. The experiments are executed on publicly available fundus datasets (EyePACS and DIARETDB1) achieving an accuracy of 89 ~ 95% with AUC, sensitivity and specificity of respectively 95 ~ 98%, 74 ~ 86%, and 93 ~ 97%, for disease level grading of DR. Whilst different CNNs produce consistent classification results, the rate of picked-up image features disagreement between models could be as high as 70%.