Abstract
Diabetic Retinopathy (DR) demands precise hemorrhage detection for early diagnosis, yet manual identification faces challenges due to hemorrhagic lesions' varied sizes, complex shapes, and color similarities to surrounding tissues, which obscure boundaries and reduce contrast. To address this, we propose SAM-ada-Res, a novel dual-encoder model integrating a pre-trained Segment Anything Model (SAM) and ResNet101. SAM captures global semantic context to distinguish ambiguous lesions from vessels, while ResNet101 extracts fine-grained details through its deep hierarchical layers. Feature maps from both encoders are fused via channel-wise concatenation, enabling the decoder to localize lesions with high precision. A lightweight Adapter fine-tunes SAM for retinal tasks without retraining its backbone, ensuring task-specific adaptation. Evaluated on three datasets (OIA-DDR, IDRiD, JYFY-HE), SAM-ada-Res outperforms state-of-the-art methods in nDice (0.6040 on JYFY-HE) and nIoU (0.4182 on IDRiD), demonstrating superior generalization and robustness. An online platform further streamlines clinical deployment, enhancing diagnostic efficiency. By synergizing SAM's generalizable vision capabilities with ResNet's localized feature extraction, SAM-ada-Res overcomes key challenges in DR hemorrhage detection, offering a robust tool for early intervention. This work bridges technical innovation and clinical practicality, advancing automated DR diagnosis.