Automated lesion segmentation is essential for DR screening, but current deep learning models often lack robustness, generating false positives in low-contrast or artifact-heavy regions. This instability largely stems from a lack of anatomical understanding. While incorporating vessel structures can guide the model, obtaining pixel-level vessel annotations for training is notoriously expensive and scarce.
To address this dilemma, the research team proposed MedFuse on 15 March 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
The framework employs a zero-shot strategy where a multimodal LLM autonomously generates precise vascular priors from raw images. These priors serve as stable anatomical anchors, guiding the segmentation network to distinguish true lesions from background noise effectively.
Experiments on the DDR and IDRID datasets demonstrate that aligning visual features with anatomical priors markedly improves the robustness of MedFuse. As a mechanism-guided solution, MedFuse achieves both high performance and data efficiency, offering a viable pathway toward reliable clinical deployment.
DOI:10.1007/s11704-025-51690-5