PulmoFoundation: AI matches pathologists, cuts biopsy review burden by 68%

A team of researchers from multiple institutions has introduced PulmoFoundation, a clinically validated foundation model for comprehensive lung pathology interpretation. Built upon the pan-cancer Virchow2 model via subspecialty-specific pretraining on ~40,000 diagnostic H&E-stained whole-slide images (WSIs), PulmoFoundation was systematically evaluated on ~26,000 WSIs covering 32 clinically relevant tasks across pre-operative, intra-operative, and post-operative care. The model accurately predicts molecular markers and patient survival while achieving clinical-grade performance on core diagnostic tasks for biopsy, frozen section, and surgical resection slides. The pretraining and evaluation were designed to address the gap between general pan-cancer models and the depth needed for subspecialty-level diagnosis.

In a registered prospective study of 1,357 patients across 11 diagnostic tasks, PulmoFoundation achieved an average AUC of 92.3%. The model demonstrated significant workflow impact: using pre-specified triage thresholds, it could reduce additional second-review burden by 68.8% for biopsies and 83.0% for frozen sections, and defer 44.5% of IHC stain orders—with positive predictive values of 1.0, 0.991, and 0.966 respectively. A crossover randomized controlled trial with eight pathologists further confirmed benefits: AI assistance improved diagnostic accuracy across 4,928 case-reader pairs (91.7% with AI vs 83.8% without), reduced median diagnostic time by 19.6%, increased diagnostic confidence by 8.7%, and improved inter-rater agreement from moderate (kappa = 0.56) to substantial (kappa = 0.76). Together, these results establish PulmoFoundation as a rigorously validated decision-support system for lung pathology.