Detection and exploration of diabetic retinopathy using advanced explainable AI (XAI) with distinctive features along with automated report generation utilizing the deep learning method

Abstract

In a world full of technologies and high screening usage, affected eyes need more attention. A serious disease that causes blindness is related to one of the most infamous one named as Diabetic retinopathy (DR). It leads to blindness if not detected early. Unfortunately, the detection process is time consuming and requires prior intensive labour. Our paper mainly focuses on how to propose something useful so that it has a framework which includes solving five class severity DR grading issues along with explainable AI (XAI) with quality-aware deep learning methods examining the fundus images. Merging two datasets one being APTOS 2019 and the other being the kaggle competitions Diabetic retinopathy dataset into a large and curated one which triggers the work of severe class-imbalance and varieties of different quality images. Our paper explores about 5 different state-to-art architectures with some notable results to have recorded for along the entire paper. The architectures listed as ConvNeXt, CoAtNet, Hybrid CoAtNet–ConvNeXt, MaxViT and Vision Mamba which were worked upon some integrated preprocessing-loss, evaluation which stands upon the clinically checkmark point. To further elaborate on the working, this paper works on lesion-level integration to build the clinical trust focusing on the impacted regions more following the predictions of the models. To summarize, our paper proposes a robust architecture which focuses on a practical way to demonstrate building a robust, clinically trustworthy and accurate framework which will make a difference in our living real world.