X-RetinaAI: Interpretable AI for Eye Disease Classification in Healthcare Environment

Integration of Convolutional Neural Networks (CNN) methodologies in the medical domain, as an application of healthcare, has seen unprecedented development in the past few years. The arrival of innovative CNN architectures in the field has provided domain experts with an additional diagnostic tool that increases their efficiency, along with providing distant healthcare support to remote regions to further develop a healthy and safe environment. Our work X-RetinaAI focuses on creating such an architecture that efficiently diagnoses patient retinal images, obtained through Optical Coherence Tomography, and classifies them into 4 different classes: Normal, Diabetic Retinopathy, Cataract, and Glaucoma, using pre-trained models such as VGG-16, and MobileNet, achieving a validation accuracy of 84.83% and 88.15% respectively. As a recent technological advancement, the Explainable Artificial Intelligence(X-AI) technique of Integrated Gradients is used to compute attribute scores and identify the features from the retinal images that have contributed most to the model's prediction. Further utilizing the features highlighted by the Integrated Gradients technique used in X-RetinaAI, a comparative analysis is performed to distinguish them from the meaningful features categorized by the ophthalmologists, so that the model can be re-trained while adding more weightage to these meaningful features, ensuring the robustness and reliability of the proposed model in classifying retinal diseases in the healthcare industry.