Towards Interpretable Foundation Models for Retinal Fundus Images
DualIFM is built to show why it makes retinal-image predictions, not just make them.
The paper proposes a foundation model for color fundus photos using a BagNet backbone, producing class evidence maps tied to the model’s decision process. It also adds a 2D projection layer during pretraining so researchers can inspect learned clinical clusters and possible spurious correlations. Trained on more than 800,000 fundus photographs, DualIFM reportedly matches RETFound-level performance with far fewer parameters and remains interpretable on out-of-distribution data. Code and pretrained models are described as available. HF Daily Papers' note
The paper proposes a foundation model for color fundus photos using a BagNet backbone, producing class evidence maps tied to the model’s decision process. It also adds a 2D projection layer during pretraining so researchers can inspect learned clinical clusters and possible spurious correlations. Trained on more than 800,000 fundus photographs, DualIFM reportedly matches RETFound-level performance with far fewer parameters and remains interpretable on out-of-distribution data. Code and pretrained models are described as available. HF Daily Papers' note
score 4