Samuel Ofosu Mensah, Camila Roa, Kerol Djoumessi, Philipp Berens
5 min
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, a critical issue in high-stakes domains such as medical imaging. We propose DualIFM, a foundation model that is interpretable-by-design via a BagNet backbone whose small receptive fields generate class evidence maps that are faithful to the model's decision-making process. Additionally, DualIFM incorporates a $2D$ projection layer during pretraining that enables direct visualization of the representation space, providing a dataset-level view of the learned structure including meaningful clinical clusters as well as potential spurious correlations. We trained DualIFM on over 800,000 color fundus photographs from various sources to learn generalizable representations for different downstream tasks. Our model achieves performance comparable to RETFound, which has $16\times$ more parameters, while providing interpretable predictions on out-of-distribution data. These results suggest that large-scale SSL pretraining paired with inherent interpretability can lead to robust representations for retinal imaging. Code and pretrained models are available at github.com/berenslab/interpretable_FM.
Alex: It's a projection layer that takes the model's internal data — which exists in a very high-dimensional space, far too complex to visualise — and compresses it down into two dimensions. Imagine taking a vast, tangled network of thousands of patient records and flattening it onto a single page so you can see which cases cluster together.
Sam: And clinicians can actually use that map to understand how the model is grouping different disease stages?
Alex: Yes. You can see whether patients with similar diagnoses are landing near each other on the map, which is a useful sanity check. It also reveals what the researchers call "spurious correlations" — cases where the model might be picking up on the wrong signal entirely. If the model is focusing on something irrelevant, it shows up visually, and a clinician can catch it.
Sam: That's a meaningful safeguard. But here's the obvious question — if the model is so much more transparent, does it pay for that with lower accuracy?
Alex: That's the critical trade-off, and the authors address it directly. Dual-IFM achieves performance comparable to much larger models — including one called RETFound, which has roughly sixteen times more parameters. More parameters generally means more computational power and more data the model can absorb. So Dual-IFM is reaching similar accuracy with a fraction of the complexity.
Sam: So it's not sacrificing performance for transparency. Is there a catch?
Alex: There is one significant limitation: memory. Because BagNet relies on large spatial feature maps to maintain that local, patch-by-patch focus, it demands more memory than you might expect given its size. That's a real constraint for hospitals with limited hardware.
Sam: So the trade-off isn't accuracy versus transparency — it's transparency versus memory footprint.
Alex: That's a precise way to put it. If a clinical setting has the hardware to support it, they get a model that produces an explicit, faithful map of its own reasoning — something you can audit, question, and explain to a patient. The researchers frame this as a meaningful step toward real clinical deployment, where both accuracy and transparency are non-negotiable.
Sam: It's a different way of thinking about trust in AI — not just "does it get the right answer," but "can we follow the logic that got it there."
Alex: And in medicine, that second question may matter just as much as the first. Thanks for listening to ResearchPod.