Accurate and Explainable Retinal Disease Recognition via DCNFIS

Mojtaba Yeganejou, Mohammad Keshmiri, Scott Dick

Abstract: The accuracy-interpretability tradeoff is a significant challenge for Explainable AI systems; if too much accuracy is lost, an explainable system might be of no actual value. We report on the ongoing development of the Deep Convolutional Neuro-Fuzzy Inference System, an XAI algorithm that has to this point demonstrated accuracy on par with existing convolutional neural networks. Our system is evaluated on the Retinal OCT dataset, in which it achieves state-of-the-art performance. Explanations for the system’s classifications based on saliency analysis of medoid elements from the fuzzy rules in the classifier component are analyzed.