Dual-stage dynamic hierarchical attention framework for saliency-aware explainable diabetic retinopathy grading

Abstract

Diabetic Retinopathy (DR) is a serious complication of diabetes, which leads to severe vision impairment if left untreated. For effective treatment, it is crucial to detect the disease early and grade it accurately. In recent years, convolutional neural networks have shown promising results in automated DR grading, yet their black-box nature challenges their interpretability. To address this, we propose a novel deep learning framework leveraging hierarchical attention refinement to dynamically highlight lesion salient regions in retinal images. The proposed model employs a Residual Network-18 backbone network to capture basic semantic feature representation from retinal fundus images, followed by a channel-wise and spatial weighed attention encoding to generate an initial saliency representation. This is further enhanced with a hierarchical cross attention mechanism to produce enriched saliency maps. Concurrently, these saliency maps guide the final decision-making process, thereby enhancing interpretability and assisting in accurate DR grading. Experimental results validate the efficacy of the proposed model in enhancing the performance of DR grading across multiple evaluation metrics. Further, quantitative and qualitative analysis of the generated saliency maps demonstrated substantial enhancements in pinpointing lesion areas within fundus images, leading to enhanced explainability and diagnostic accuracy of the model’s predictions.