PulmoFoundation: a randomized crossover trial confirms the gain of AI in pulmonary pathology

With assistance, the diagnostic accuracy of eight pathologists increased from 83.2% to 91.7%, and the median analysis time fell by 18.3%.

Pathological analysis guides lung cancer diagnosis, treatment selection, and prognosis, but current approaches rely on specialized models, each dedicated to an isolated goal. Pan-cancer foundation models offer versatility without subspecialty depth, and have not previously been prospectively validated in real-world conditions.

PulmoFoundation was built on Virchow2 through subspecialty pre-training of approximately 40,000 whole hematoxylin-eosin-stained slides, and then evaluated on nearly 26,000 slides and 32 clinically relevant tasks, from biopsy collection to surgical specimen to extemporaneous examination. In a prospective registered study of 1,357 patients and 11 diagnostic tasks, the mean AUC was 92.3%.

A crossover trial in eight pathologists

In addition to prospective validation, the authors conducted a randomized controlled crossover trial with eight pathologists, i.e. 5,264 case-reader pairs: diagnostic accuracy increased from 83.2% without assistance to 91.7% with. The median time to diagnosis decreased by 18.3%, reported confidence increased by 9%, and inter-observer agreement increased from moderate to substantial (kappa from 0.55 to 0.76). With predefined sorting cut-offs, the model would avoid a second reading for 68.8% of biopsies and 83% of extemporaneous examinations, and defer 44.5% of immunohistochemistry requests.