Interpretable DualIFM Foundation Model for Retinal Imaging
August 3, 2026
DualIFM is an interpretable-by-design foundation model for retinal fundus images using a BagNet backbone. It generates class evidence maps and utilizes a 2D projection layer to visualize clinical clusters and identify spurious correlations in the representation space.
HOW THIS AFFECTS YOU
●
researcherThe architecture allows for direct visualization of learned structures in medical SSL models.
●
healthThis increases trust in AI-driven retinal diagnostics by providing visual evidence for decisions.