Visualizing the Indicators of Diabetic Retinopathy Learnt by Convolutional Neural Networks

This study proposes a novel application of visualizing features learnt by convolutional neural networks with the aim to further the understanding of Diabetic Retinopathy. A convolutional neural network is first trained to recognize and classify fundus images of diabetic and non-diabetic patients. The network is then visualized, using a technique of pixel optimization, to discover the features that the trained network looks for to classify the image. Through this novel application of network visualization, we show that critical features for diabetic retinopathy can be re-discovered, leaving great scope for its application in scarcely explored diseases using minimal resources.