Visualizing and Understanding Inherent Image Features in CNN-based Glaucoma Detection

Convolutional neural network (CNN)-based methods have achieved state-of-the-art performance in glaucoma detection. Despite this, these methods are often criticized for offering no opportunity to understand how classification decisions are made. In this paper, we develop an innovative visualization strategy that allows the inherent image features contributing to glaucoma detection at different CNN layers to be understood. We also develop a set of interpretable notions to better comprehend the contributing image features involved in the disease detection process. Extensive experiments are conducted on publicly available glaucoma datasets. Results show that the optic cup is the most influential ocular component for glaucoma detection (overall Intersection over Union (IoU) score of 0.18), followed by the neuro-retinal rim (NR) with IoU score 0.17. With an overall IoU score of 0.16 vessels in the photograph also play a considerable role in the disease detection.