🤖 AI Summary
To address weak fine-feature perception, strong background interference, and model redundancy in diabetic retinopathy microlesion detection, this paper proposes a lightweight yet high-accuracy YOLOv8n-based framework. Methodologically, it introduces (1) dynamic KWConv and the C2f-KW module to enhance local sensitivity to subtle lesions; (2) a Feature-Focusing Diffusion Pyramid Network (FDPN) for adaptive multi-scale contextual feature fusion; and (3) a lightweight Shared Detection Head (GSDHead) to drastically reduce parameter count. Experimental results demonstrate that the proposed model reduces parameters by 20.7% and improves mAP@0.5 by 4.1% and recall by 7.9% over baseline YOLOv8n. Moreover, it outperforms mainstream one-stage detectors—including YOLOv5n and YOLOv10n—in both accuracy and efficiency, achieving an optimal balance between computational cost and detection performance for clinical-grade microlesion identification.
📝 Abstract
Early detection and diagnosis of diabetic retinopathy is one of the current research focuses in ophthalmology. However, due to the subtle features of micro-lesions and their susceptibility to background interference, ex-isting detection methods still face many challenges in terms of accuracy and robustness. To address these issues, a lightweight and high-precision detection model based on the improved YOLOv8n, named YOLO-KFG, is proposed. Firstly, a new dynamic convolution KWConv and C2f-KW module are designed to improve the backbone network, enhancing the model's ability to perceive micro-lesions. Secondly, a fea-ture-focused diffusion pyramid network FDPN is designed to fully integrate multi-scale context information, further improving the model's ability to perceive micro-lesions. Finally, a lightweight shared detection head GSDHead is designed to reduce the model's parameter count, making it more deployable on re-source-constrained devices. Experimental results show that compared with the base model YOLOv8n, the improved model reduces the parameter count by 20.7%, increases mAP@0.5 by 4.1%, and improves the recall rate by 7.9%. Compared with single-stage mainstream algorithms such as YOLOv5n and YOLOv10n, YOLO-KFG demonstrates significant advantages in both detection accuracy and efficiency.