An interpretable dual attention network for diabetic retinopathy grading: IDANet

Amit Bhati, Neha Gour, Pritee Khanna, Aparajita Ojha, Naoufel Werghi

Abstract: Highlights•A dual attention model for fine-grained DR-grading.•A self-attention based BSA block to learn lesion-specific contextual information for the entire pixel space.•A CPA block to minimize redundancy and focus on lesion-specific details.•Model interpretability is validated on the IDRiD dataset using the LIME method.•Validation of the proposed model using statistical and runtime performance analysis.