Explainable and Uncertainty-Aware Diabetic Retinopathy Detection Using DRN-CBAM-GBAW for Lesion-Based Interpretability

Diabetic Retinopathy (DR) is a major cause of vision impairment worldwide, arising from microvascular damage in the retina due to prolonged hyperglycemia. Early identification of retinal lesions such as microaneurysms, hemorrhages, and exudates is essential to prevent disease progression. However, existing Dilated Residual Networks (DRN) with Region of Interest (ROI) modules primarily focus on lesion detection without addressing model uncertainty or interpretability, limiting their clinical applicability. To overcome these limitations, this research proposes a hybrid Dilated Residual Network (DRN) backbone with a Convolutional Block Attention Module (CBAM) integrated with a Gaussian-Based Adaptive Weighting (GBAW). The proposed DRN-GCA-Net model is an explainable and uncertainty-aware framework for lesion-based DR detection. It generates spatially continuous, uncertainty-informed attention maps that enhance lesion localization and provide transparent associations between predictions and pathological regions. In addition, predictive uncertainty is quantified to support reliable refer versus no-refer clinical decision-making. The proposed DRN-GCA-Net model is evaluated on benchmark DiaRetDB1 dataset and IDRiD dataset, which contain high-resolution, expert-annotated fundus images. Experimental results demonstrate that the DRN-GCA-Net achieves superior lesion interpretability, localization precision, and diagnostic reliability attaining the accuracies of$\text{9 9. 5 \%}$and$\text{9 9. 4 2 \%}$, respectively.