Retinopathy Lesions Segmentation Using YOLOv9 and Grad-CAM: An Advanced Approach for Disease Grading and Localization

Abstract

Diabetic retinopathy (DR) is a leading cause of blindness, and accurate detection and segmentation of retinal lesions are critical for effective diagnosis and treatment. This paper presents an automated approach for DR lesion segmentation using the YOLOv9 model, combined with Grad-CAM for interpretability. YOLOv9 was chosen for its advanced object detection capabilities, enabling effective localization of small lesions, while Grad-CAM provides heatmaps for better interpretability, making it easier to understand the model's focus areas during prediction. The proposed method was evaluated on two public datasets, DDR and IDRiD, achieving high segmentation accuracy across various lesion types, including microaneurysms (MAs), hemorrhages (HE), exudates (EX), and soft exudates (SE). For DDR, the model achieved mean Average Precision (mAP) scores of 0.769 for MAs, 0.917 for HEs, 0.843 for EX, and 0.960 for SE at an IoU threshold of 0.5, with an overall mAP of 0.872 across all classes. The precision-recall curve and normalized confusion matrix further demonstrated the model's high recall and precision, especially for hemorrhages and soft exudates, while microaneurysms showed slightly lower precision due to background misclassifications. This approach combines high detection accuracy with interpretability, making it a promising tool for aiding ophthalmologists in DR diagnosis. Future work will focus on refining lesion detection for challenging classes and exploring multi-modal data integration for broader retinal disease applications.