With the rapid increase in the number of diabetes patients worldwide, diabetic retinopathy (DR) has become one of the leading complications threatening patients’ vision health. Early and accurate diagnosis is crucial for effective treatment and prevention of vision loss. Currently, the diagnosis of DR primarily relies on fundus examination, which is not widely accessible. However, the emergence of automated diagnostic methods has shed light on this problem, although it still faces challenges in localizing and classifying complex pathological areas, limited accuracy in detecting subtle abnormalities. To address these limitations, we propose an innovative DR image classification method that enhances the focus on lesion region and improves the differentiation of DR images. Firstly, we adopt a few-shot learning strategy to precisely segment key lesion areas, including microaneurysms, hemorrhages, hard exudates, soft exudates, and vessels with limited annotated data. Secondly, we design a dual-path network architecture that employs a backbone to extract global features from the original fundus images and a lightweight auxiliary network to extract lesion-specific features from the segmentation maps. Finally, we introduce a lesion-guided attention mechanism based on the Convolutional Block Attention Module to fuse the dual-path features and guide the backbone network to focus on critical pathological regions. Experimental results demonstrate that our model achieves state-of-the-art accuracy in DR classification, with accuracies of 87.18% on the ATOPS dataset and 85.93% on the DDR dataset. These results underscore its strong generalization ability and efficacy as a solution for automated DR diagnosis.
Ying Shao (Fri,) studied this question.