A pathology slide is a gigapixel image — often 100,000 pixels on a side — and expert labels are scarce. So the field did what worked for text and cells: pretrain self-supervised on millions of unlabeled slides. UNI, Virchow and Prov-GigaPath learn general histology representations at staggering scale (Prov-GigaPath alone: 1.3 billion image tiles from 171,189 whole slides), CONCH adds language so you can query a slide in words, and rare-cancer detection starts to work where labels are thinnest. But the same year’s honest reckoning is bracing: these models still encode the hospital, not just the biology — stain and scanner ‘site signatures’ that bias predictions and can turn a benchmark win into a batch effect. It’s imaging-times-AI at clinical scale, and the prove-it discipline the lab keeps returning to, at its sharpest.
