Trustworthy AI · computational pathology

From saliency maps to biological concepts

An interface concept for making glioma pathology foundation models interpretable — explaining a prediction through the tumour features a pathologist names, and showing which explanations actually hold up.

Synthetic tissue · transparent stand-in model · not for clinical use

A pathology foundation model reads a gigapixel slide and returns a glioma diagnosis, but not why. The usual answer is a saliency map over the image. This page contrasts that with a concept-based explanation: the same prediction broken into named morphological features — cellular density, nuclear atypia, mitotic activity, microvascular proliferation, necrosis. Use the controls to switch views, test how each explanation behaves under a tiny input perturbation, and search for biologically similar regions.

Slide region · 6 × 6 tiles click a tile to inspect & find similar
concept attribution saliency selected similar region

Model output

aggregated over all tiles

0.00 model confidence
Predictive uncertainty0.00

Why — concept contributions

Click a concept to highlight where it appears on the slide.

Stability under input perturbation

Saliency map
Concept explanation

Toggle Inject input noise to perturb the input and measure how much each explanation changes. Smaller is better.

Selected region

No tile selected. Click a tile on the slide to see its concept profile.

Content-based retrieval

Pick any tile; the model ranks the rest by similarity in concept space and returns the closest regions — a starting point for "show me more tissue that looks like this".

Select a tile above to retrieve its most similar regions.