Hong Kong University of Science and Technology / Southern Medical University / Guangdong Provincial Key Laboratory of Molecular Tumor Pathology / Fourth Military Medical University / University of Science and Technology of China / Zhejiang University / HaploX Biotechnology / Hebei Medical University / 900th Hospital of the PLA Joint Logistic Support Force / Shandong Provincial Qianfoshan Hospital
A subspecialty lung-pathology foundation model, fine-tuned from Virchow2 and prospectively validated across 32 clinical tasks spanning the lung diagnostic workflow.
PulmoFoundation is a subspecialty foundation model for lung pathology, developed by a consortium led by the Hong Kong University of Science and Technology (with senior author Hao Chen) alongside multiple Chinese medical centers and posted to arXiv in May 2026. Rather than building a general-purpose pathology model from scratch, the authors adapt an existing whole-slide-image foundation model to a single organ system, arguing that subspecialty focus is what enables clinically deployable performance across the full diagnostic workflow.
The model addresses a practical gap in computational pathology: general foundation models learn broad morphological representations but are rarely validated end-to-end against the actual decisions a lung pathologist makes, which span pre-operative biopsies, intra-operative frozen sections, and post-operative resection specimens. PulmoFoundation is built by subspecialty-specific pretraining of Virchow2 — a self-supervised vision transformer from Paige (already cataloged here as part of the Virchow family) — on roughly 40,000 diagnostic hematoxylin-and-eosin whole-slide images, then evaluated on 32 clinically relevant tasks covering the lung-pathology pipeline.
What distinguishes the work is the rigor of its clinical validation. Beyond retrospective benchmarks, the authors report a prospective study across 1,357 patients and a randomized reader study measuring how AI assistance changes pathologist behavior, making PulmoFoundation one of the more thoroughly validated organ-specific pathology models reported to date.
PulmoFoundation inherits a Vision Transformer backbone from Virchow2, which uses a ViT-H/14 architecture (632 million parameters) pretrained with a DINOv2-derived self-supervised objective on millions of whole-slide images. The PulmoFoundation work adds a subspecialty pretraining stage on approximately 40,000 lung diagnostic H&E WSIs to adapt these representations to thoracic tissue, then attaches task-specific heads for the downstream classification and decision tasks. Evaluation spans roughly 26,000 WSIs across 32 tasks. On the prospective cohort of 1,357 patients (11 tasks), the model reached an average AUC of 92.3%. The randomized reader study covered 4,928 case-reader pairs from 8 pathologists, where AI-assisted reads improved accuracy to 91.7% (versus 83.8% unassisted), reduced median diagnostic time by 19.6%, increased diagnostic confidence by 8.7%, and raised inter-rater agreement to kappa = 0.76. The work is a preprint (arXiv eess.IV / cs.CV), so these figures await peer review.
PulmoFoundation targets the lung-pathology diagnostic workflow directly: triaging biopsies, supporting intra-operative frozen-section reads, and assisting post-operative resection assessment. Its reported ability to flag cases for second review and to defer immunohistochemistry orders suggests use as a workflow-prioritization and resource-allocation tool, potentially reducing turnaround time and ancillary-test costs in high-volume thoracic pathology services. The reader-study results — faster reads, higher confidence, and better inter-rater agreement — position it as a decision-support aid intended to augment pathologists rather than replace them, of particular value in settings with limited subspecialty thoracic expertise.
PulmoFoundation illustrates a growing strategy in computational pathology: rather than scaling ever-larger general foundation models, adapt an existing strong backbone (here, Virchow2) to a single organ system and validate it against real clinical decisions. Its prospective and randomized evaluations go beyond the retrospective benchmarks common in the field, providing direct evidence that foundation-model assistance can change pathologist accuracy, speed, and agreement. Key limitations are that the results are from a preprint pending peer review, the validation cohorts are drawn from collaborating Chinese medical centers, and — at the time of release — no model weights or code repository have been made publicly available, which constrains independent reproduction and external deployment.