Gradient-based Explainable Artificial Intelligence Methods for Eye Disease Classification

Abstract

The use of artificial neural networks in eye disease recognition holds great promise. However, with the increasing complexity of neural network architectures and the inability to trace decision-making pathways, trust in the technology is declining. The use of explainable artificial intelligence in eye disease recognition will provide confidence to ophthalmologists. Methods based on gradient calculation (gradient-based) are becoming popular for image processing tasks. Our study presents a theoretical description and practical implementation of several explanatory artificial intelligence methods (Grad-CAM, Vanilla Gradient, Guided IG, Blur IG, XRAI) on the example of explanatory artificial neural network results for the task of ophthalmic disease image classification (diabetic retinopathy, cataracts, glaucoma).