Automatic grading diabetic retinopathy in color fundus image: Cascaded hybrid attention network

Abstract

Deep learning techniques have been introduced to automatically assess Diabetic Retinopathy (DR), a prevalent vision-impairing condition, aiding ophthalmologists in formulating personalized treatment strategies for patients. However, most Artificial Intelligence (AI) for grading DR are constrained by their limited capacity to discern lesion details or necessitate manual lesion labeling, leading to suboptimal assessment outcomes or increased workload. To enhance both efficiency and precision in grading, this paper proposes a Cascaded Hybrid Attention Network (CHAN) for DR grading. Residual Hybrid Attention Modules (RHAM) is designed to extract features, incorporating several cascaded Hybrid Attention Module (HAM) and a convolution layer with a residual connection. The approach integrates channel attention with multi-head self-attention, leveraging their combined strengths: global statistical analysis and robust local adaptation. The grading score is derived by integrating DR features across various layers, from shallow to deep. The CHAN model, tested on the MESSIDOR dataset, demonstrates superior accuracy in DR grading, consistently surpassing current leading methods across evaluation metrics.