UNET++ AND EXPLAINABLE AI FOR EARLY DETECTION OF DIABETIC RETINOPATHY FROM FUNDUS IMAGES

Abstract

One of the most prevalent causes of vision loss and vision impairment among diabetic patients worldwide that can be prevented is Diabetic Retinopathy (DR). This is especially important to identify the abnormality in the retina as soon as and as accurately as possible so that it can be treated clinically and the disease controlled. In this study, an integrated framework is proposed which combines UNet++ deep learning architecture and Explainable Artificial Intelligence (XAI), to be used for early detection of DR from retinal fundus image. The proposed system will involve UNet++ to accurately segment the retinal lesion including microaneurysm, hemorrhages and exudates on the retina to facilitate better localization and characterization of the features related to the disease. A robust and complete preprocessing pipeline is performed to enhance the images quality assessment, normalization, contrast enhancement and data augmentation are performed to enhance the model robustness and generalization. To enable these models to be explained in the clinical domain and bring transparency to these models, XAI methods are integrated into the pipeline such as the visualization of Grad-CAM and saliency maps of the image regions that justify the diagnosis. Evaluation is done with publicly available annotated fundus image datasets, and using standard metrics such as accuracy, precision, recall, F1 score, Dice coefficient, and Intersection-over-Union (IoU). Experimental results show that the proposed approach can obtain high accuracy in lesion segmentation and also give meaningful explanation to images based on visual results to assist clinical decision making. UNet++ combined with explainability mechanism enhances diagnostic accuracy and build trust among clinicians. The proposed framework provides a practical and trustworthy CADe system for the large scale DR screening and early detection of DR in healthcare settings.