CPathOGen produces controlled simulated tissue samples to test what pathological AI systems are actually looking at

This diffusion model generates modified H&E-stained slides, altering one variable at a time, to assess whether an AI responds to biological factors or simply to variations in staining.

Digital pathology models predict diagnoses or survival from histological slides, but it remains unclear which tissue signals actually guide their decisions. The usual disturbances, applied to pixels or internal representations, often produce unrealistic tissue that distorts interpretation.

Verified counterfactuals

Samarth Singhal and Varang Rai propose CPathOGen, a conditional latent diffusion model that generates pairs of hematoxylin-eosin-stained images by separately controlling the spatial organization of cells, the morphology of nuclei, and the appearance of staining. The generated cells respect the requested abundance and position, and morphology and color measurements vary regularly with settings.

These interventions are then used to probe pathology encoders, classifiers, and survival models. The authors introduce an indicator, the Biology-Nuisance Sensitivity Ratio, which compares a model's sensitivity to biological factors to its sensitivity to coloration variations, which are of no clinical interest. Code and weight are published on GitHub and Hugging Face.