DualIFM: Interpretable Foundation Model for Retinal Images Matches RETFound at 1/16 Size
UniTuebingen · hf · 2026-08-10
A team from the University of Tübingen introduced DualIFM, a foundation model designed to be interpretable-by-design for retinal fundus images. This addresses the critical lack of interpretability in existing self-supervised learning (SSL) models used in high-stakes medical domains.
Core Mechanisms
- BagNet Backbone: Utilizes a BagNet architecture with small receptive fields to generate class evidence maps faithful to the model's decision-making process.
- 2D Projection Layer: Incorporates a 2D projection layer during pretraining to enable direct visualization of the representation space, revealing dataset-level structures like clinical clusters and spurious correlations.
Performance
Trained on over 800,000 color fundus photographs, DualIFM achieves performance comparable to RETFound (which has 16 times more parameters) while providing interpretable predictions on out-of-distribution data. Code and pretrained models are open-sourced.
More from Research
- Research: Unconditional Prediction Accuracy Isn't Always the Right Objective in AI Decision Processes — joshgans · 2026-08-10
- Synthetic Query Probing: Comparing Similarity Spaces Across Embeddings — pppeer · 2026-08-10
- Gödel, Escher, Bach Predicted AI: Gradient Descent as the Ultimate 'Strange Loop' — AymericRoucher · 2026-08-10
- OneEmo: A Unified Multimodal Reasoning Model for Emotion Perception — Jiahao Huang · 2026-08-10
- Making Knowledge Distillation Cheap Enough to Run at Scale — Hugging Face Blog · 2026-08-10
- DeepMind Paper Claims Current RAG Architectures Have Uncrossable Mathematical Limits — solyarisoftware · 2026-08-10