[2603.18846] Towards Interpretable Foundation Models for Retinal Fundus Images

Computer Science > Computer Vision and Pattern Recognition arXiv:2603.18846 (cs) [Submitted on 19 Mar 2026 (v1), last revised 14 Apr 2026 (this version, v2)] Title:Towards Interpretable Foundation Models for Retinal Fundus Images Authors:Samuel Ofosu Mensah, Camila Roa, Kerol Djoumessi, Philipp Berens View a PDF of the paper titled Towards Interpretable Foundation Models for Retinal Fundus Images, by Samuel Ofosu Mensah and 3 other authors View PDF HTML (experimental) Abstract:Foundation models are used to extract transferable representations from large amounts of unlabeled data, typically via self-supervised learning (SSL). However, many of these models rely on architectures that offer limited interpretability, which is a critical issue in high-stakes domains such as medical imaging. We propose Dual-IFM, a foundation model that is interpretable-by-design in two ways: First, it provides local interpretability for individual images through class evidence maps that are faithful to the decision-making process. Second, it provides global interpretability for entire datasets through a 2D projection layer that allows for direct visualization of the model's representation space. We trained our model on over 800,000 color fundus photography from various sources to learn generalizable, interpretable representations for different downstream tasks. Our results show that our model reaches a performance range similar to that of state-of-the-art foundation models with up to $16\times$ t...