Integrating Explainability in AI for Retinal Imaging: Enhancing Diabetic Retinopathy Diagnosis Accuracy

Diabetic retinopathy (DR: preventable cause of blindness till date as it is irreversible till now.) Widely available adoption of clinical features is hard because traditional models have not been interpretable; AI (Artificial Intelligence) has so far made an exception in automating diagnosis for diabetic retinopathy. Explainable AI enables the generation of raw, simple results by machines which one as a human might agree on and hence gets re-empowers physician. In this paper, Explainable AI framework is discussed in detail. This is a novel approach to deal with the diabetic retinopathy which deals with various issues such as low-labelled-data requirement, lack of validations among various populations and sub-optimal performance in real and unseen data. The paper presents the framework that includes scalable models for resource-constrained environments, multistage detection network and differential-patient diagnosis. The framework is also generalize with the latest interpretability frameworks (i.e. Grad-CAM and LIME) where explanation can be given to clinically grey matter while keeping accuracy. Clinical multi-centre research done shown that near 94.5% accuracy versus current 88% solutions currently using the framework can lead to substantial improvements in the diagnosis from the framework. Furthermore proof-of-concept studies in a smaller setting have shown that it is scalable and physician surveys found 40% increase in trust so too. The vast and critical knowledge gaps are filled by the presented work, allowing for solution to enhance the diagnostic scalability accuracy and transparency. These results explain the support of explainable AI towards the retinal imaging forward, especially at diabetic retinopathy diagnosis and will towards new breakthroughs in clinical predictive AI.