Lesion-aware generalizable deep learning framework for retinal lesion classification in diabetic retinopathy

Diabetic Retinopathy (DR) diagnosis relies on the accurate identification of key retinal lesions. However, existing deep learning models often cannot generalize across datasets due to variations in imaging conditions and the model being trained with full fundus images that contain irrelevant background information. This leads to reduced accuracy, limited robustness, and poor clinical interpretability. This study proposes a Lesion-Aware Region of Interest (RoI)-based deep learning framework, termed LARNet, for DR lesion classification. A U-Net architecture with an EfficientNet backbone is employed to segment lesion regions and produce accurate lesion region masks, isolating clinically relevant structures. These lesion RoIs are applied on the original fundus images and fused with EfficientNetV2-based global features to generate small, lesion-specific embeddings, which are then used to classify hemorrhages (HE), hard exudates (EX), soft exudates (SE), and microaneurysms (MA). In addition to high diagnostic performance, LARNet achieves high computational efficiency with fewer trainable parameters and less inference time than conventional deep learning architectures, enabling deployment in real-time clinical screening systems with limited computational capability. Cross-dataset and XAI analysis further demonstrate LARNet’s robustness and provide clinically meaningful visual explanations of predictions. Thus, our model confirms its usefulness in targeting the specific areas of lesions and facilitate the efficient early-stage detection of DR in the real-world screening environments.