An explainable deep learning approach for automated detection and grading of diabetic retinopathy from fundus images

Diabetic Retinopathy (DR) is a leading cause of preventable blindness, and it is important to accurately detect it as early as possible, as well as to measure the severity to provide early clinical indicators. Fundus image assessment in the manual mode is subjective and labor-intensive, and is hard to scale, which has encouraged automated approaches. Current deep learning methods are usually based on either convolutional neural networks (CNNs) or transformer-based ones, focusing on local lesion features or global retina context separately. In addition, relational dependencies among lesions that are clinically significant in the case of severity development are under-modeled. This paper introduces a relational, hybrid deep learning model that integrates convolutional, transformer-based, and graph attention models to classify binary and multiclass DR. A ResNet-Graph Attention Network (ResNet-GAT) is introduced to explicitly model spatial relational dependencies among regional feature descriptors and iscompared with independent stand-alone ResNet-50 and Swin Transformer models. Also, a hybrid ResNet-Swin Transformer model combines fine-grained lesion representations and hierarchical global context. Experiments with the APTOS 2019 data show that the hybrid model has an accuracy of 98.09% with a Quadratic Weighted Kappa (QWK) of 0.9618 when used to classify binary, and 94.89% with a QWK of 0.9726 when used to classify five classes of severity. Robustness is proven by cross-dataset inference on IDRiD. The interpretability analysis performed through Grad-CAM shows that the predictions are made based on clinically significant areas, such as microaneurysms, hemorrhages, and exudates. It is demonstrated that performance and interpretability are enhanced by a combination of local, global, and relational representations in automated screening of DR.