Autonomous learning of pathologists' cancer grading rules
Deep learning reveals that tissue morphology contains rich pathophysiological information beyond human understanding. However, approaches to convert these spatially distributed signals into subcellular insights informing disease mechanisms are lacking. We introduce Delta-Marches, an interpretability-first approach that nominates distinguishing morphological features rather than explaining existing models' decisions. Delta-Marches simulates idealized morphological changes between classes by coupling latent-space traversals with generative AI. Comparing each image to its class-shifted counterpart allows feature extractors to infer the most affected aspects, reducing sample-to-sample variability and resolving transformations at subcellular resolution. Prototyped in renal carcinoma grading, Delta-Marches generates realistic grade transitions and pinpoints tumor-cell nuclear phenotypes as key determinants. It also reveals reduced vasculature with increasing grade, a known pattern absent from standard rubrics. Applied to dysplastic progression in colorectal tissue, it recovers glandular remodeling and goblet-cell loss, generalizing across tissue types and scales. These results show Delta-Marches parses complex image phenotypes and catalyzes hypothesis generation.