OBJECTIVE

To enhance the automatic detection precision of diabetic retinopathy (DR) lesions, this study introduces an improved YOLOv8 model specifically designed for the precise identification of DR lesions.

METHOD

This study integrated two attention mechanisms, convolutional exponential moving average (convEMA) and convolutional simple attention module (convSimAM), into the backbone of the YOLOv8 model. A dataset consisting of 3,388 ultra-widefield (UWF) fundus images obtained from patients with DR, each with a resolution of 2,600 × 2048 pixels, was utilized for both training and testing purposes. The performances of the three models-YOLOv8, YOLOv8+ convEMA, and YOLOv8+ convSimAM-were systematically compared.

RESULTS

A comparative analysis of the three models revealed that the original YOLOv8 model suffers from missed detection issues, achieving a precision of 0.815 for hemorrhage spot detection. YOLOv8+ convEMA improved hemorrhage detection precision to 0.906, while YOLOv8+ convSimAM achieved the highest value of 0.910, demonstrating the enhanced sensitivity of spatial attention. The proposed model also maintained comparable precision in detecting hard exudates while improving recall to 0.804. It demonstrated the best performance in detecting cotton wool spots and the epiretinal membrane. Overall, the proposed method provides a fine-tuned model specialized in subtle lesion detection, providing an improved solution for DR lesion assessment.

CONCLUSION

In this study, we proposed two attention-augmented YOLOv8 models-YOLOv8+ convEMA and YOLOv8+ convSimAM-for the automated detection of DR lesions in UWF fundus images. Both models outperformed the baseline YOLOv8 in terms of detection precision, average precision, and recall. Among them, YOLOv8+ convSimAM achieved the most balanced and accurate results across multiple lesion types, demonstrating an enhanced capability to detect small, low-contrast, and structurally complex features. These findings support the effectiveness of lightweight attention mechanisms in optimizing deep learning models for high-precision DR lesion detection.