Abstract: The small size and dense distribution of bleeding point lesions in the fundus image of diabetic retinopathy make it difficult for the existing algorithms to achieve accurate detection and localization of the lesions. A RCA-YOLO bleeding lesion detection algorithm is proposed. Based on YOLOv5s, the RCA-Net module is first used in the backbone network, so that the network can obtain the connection between each channel while retaining the location information of the target lesion, and enhance the feature extraction and localization ability of the network for the bleeding area. In the feature fusion stage, the lightweight feature pyramid network Tiny-BiFPN is adopted to reduce the number of network parameters and achieve high-efficiency multi-scale feature fusion. Finally, a small target feature enhancement module is proposed to improve the detection accuracy of the algorithm for small bleeding point lesions. The experimental results show that the improved RCA-YOLO algorithm can accurately detect and locate bleeding point lesions, and the average detection accuracy (mAP) can reach 79.3%, which is 9.5 percentage points higher than that of YOLOv5s algorithm, and its detection results are also better than mainstream algorithms such as Faster R-CNN, YOLOv6s, YOLOv7 and YOLOv8s.

Key words: fundus image, diabetic retinopathy, object detection, hemorrhagic lesion, feature fusion