XFM: An Explainable Knowledge-Fused Vision Foundation Model for Improving Clinical Diagnosis of Low-Prevalence...

Foundation models in ophthalmology, often pre-trained on extensive datasets, exhibit exceptional generalization and emergent capabilities that are absent in smaller-scale specialized models. This study first investigated the adaptation of ophthalmic foundation models to detect low-prevalence retinal diseases in real-world clinical settings with low-data regimes. We then bridged the gap in exploring the use of fine-grained prior-knowledge infusion and SAM-guided cycle constraint regularization to enhance the explainability of the foundation models from both qualitative and quantitative perspectives. Evaluated on two newly constructed public datasets (FundusData-FS and OTFID), our foundation model-based solution demonstrates effective transfer learning and few-shot learning fine-tuning performance for multiple low-prevalence retinal diseases. Our experiments demonstrate that prior-knowledge infusion and SAM-guided regularization enhance both the performance and the explainability of the foundation model. For example, our method achieves over 9% accuracy improvement and superior AUC performance (an 8% gain over RETFound) in GradCAM-positive perturbation testing, with statistically significant improvements (p <0.05) in the FundusData-FS dataset. These findings highlight the potential of explainable ophthalmic foundation models for trustworthy AI in clinical practice.