Breast pathology foundation model LoRA-adapted from Virchow2 on 57,271 breast slides, validated prospectively for biopsy and frozen-section triage.
No providers recorded yet. Browse all providers
A breast pathologist handles three very different kinds of glass. A core-needle biopsy offers a few millimetres of tissue from which to call malignancy and infer receptor status. A frozen section must be read in minutes, through ice-crystal artefacts, while the patient is still on the table. A resection supports staging, grading and prognosis. Pan-cancer pathology foundation models learn broad morphology from millions of slides across many organs, but breast-specific distinctions, such as lymphovascular invasion or the subtle cues that separate a metastasis from reactive tissue, make up only a small share of what they see.
BRAVE (BReAst Vision pathology foundation modEl) addresses this with breast-adaptive pretraining. Rather than training from scratch, it starts from the pan-cancer encoder Virchow2 (part of the Virchow family) and continues self-supervised learning on 57,271 breast whole-slide images, updating only low-rank adapter weights so the general knowledge in the frozen backbone is preserved. The result is a single frozen patch encoder that is reused unchanged across every downstream task.
BRAVE was developed by Hao Chen's group at the Hong Kong University of Science and Technology with pathology departments across mainland China and Hong Kong, and posted to arXiv in May 2026. It follows the same subspecialty-adaptation strategy as the group's lung model PulmoFoundation and frozen-section model CRISP.
BRAVE uses a ViT-H/14 architecture (32 layers, hidden size 1,280, 16 heads) that operates on 224 × 224 pixel tiles at 20× magnification. Its 2,560-dimensional embedding concatenates the class token with the mean of the patch tokens. Adaptive pretraining used a DINO teacher–student objective on more than 113 million patches from 57,271 breast slides drawn from 15 sources in North America, Europe and Asia. Training ran for 882,000 iterations on eight H800 GPUs. The study's full resource is 101,638 slides from 32 sources, with 44,367 of those slides reserved for downstream evaluation. For slide- and case-level tasks, an attention-based multiple-instance learning (ABMIL) aggregator and a linear head are trained on frozen BRAVE features.
Across 75 retrospective cohorts, BRAVE had a mean rank of 1.43 against Virchow2 (2.57), UNI (2.99), CONCH (3.43) and PLIP (4.58). It beat Virchow2 with statistical significance on 27 of 30 classification tasks. The largest gains were in metastasis detection, T-stage prediction, mutation prediction, response to neoadjuvant chemotherapy and frozen-section margin assessment. In prospective validation at three centres, BRAVE excluded 76.9% of negative biopsies (NPV 0.953) and 70.1% of negative frozen sections (NPV 0.973), and it triaged 78.8% of post-operative subtyping cases as clear-cut (NPV 1.000). In crossover reader studies, AI assistance raised average balanced accuracy from 88.5% to 95.1% (OR 3.14). The adjusted hazard ratios were 4.79 for disease-free survival and 8.14 for overall survival.
BRAVE is aimed at breast pathology services. It can rule out clearly negative biopsies and frozen sections from routine review, give intra-operative reads a second opinion under time pressure, rescue missed Ki67-positive cases through AI-assisted second review, and prioritize resections for molecular testing. Researchers can also use it as a frozen breast-specific feature extractor for new multiple-instance learning tasks. The code covers LoRA pretraining, ABMIL classification and survival modelling.
BRAVE is one of the first organ-specific pathology foundation models for breast cancer whose evidence runs from retrospective benchmarks through locked-threshold prospective validation to human–AI reader studies. It supports the view that inexpensive adapter-based specialization of a strong pan-cancer backbone pays off most on fine-grained, organ-specific tasks. The work is a preprint awaiting peer review, and all of its prospective cohorts come from Chinese centres. The training and evaluation code is public, but the weights are available only on request to the authors, with public release deferred until the paper is accepted. The model and code are released under CC-BY-NC-ND 4.0, which limits use to non-commercial research.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.