A new AI model can paste convincing fake tissue lesions into pathology slides — and expert pathologists can barely tell the difference.
Researchers introduced PathoGen, a diffusion-based model that inpaints synthetic lesions into benign histopathology images with enough fidelity to fool six expert pathologists 42.25% of the time (distinguishing real from fake landed at only 57.75% accuracy, barely above a coin flip). The system was validated across kidney, skin, breast, and prostate tissue datasets. When synthetic images were added to training sets, AI segmentation models improved by up to 0.18 Dice score — a meaningful jump — with the biggest gains in exactly the scenarios where real data is hardest to come by.
Annotated pathology data is one of the nastiest bottlenecks in medical AI: rare cancers may have only dozens of confirmed cases globally, and labeling a single slide requires a specialist. PathoGen sidesteps both problems by generating not just realistic images but pixel-level annotations alongside them, cutting out the most expensive part of the pipeline.
The broader pattern here is familiar — synthetic data rescuing AI where real data is thin — but histopathology is harder than most domains because tissue architecture is structurally complex and pathologists are trained to notice subtle morphological tells. That PathoGen can slip past that scrutiny most of the time is the more interesting result than any benchmark number.