Breast-specialized multimodal pathology foundation model for core needle biopsy diagnosis, with conformal risk control gating report release.
No providers recorded yet. Browse all providers
Breast core needle biopsy (CNB) is the first diagnostic step for most suspected breast cancers, and one of the harder specimens to read. The sampled tissue is small, lesions are heterogeneous, and diagnostically adjacent entities overlap morphologically, so a report has to be built up in stages — cancer versus noncancer, then in situ versus invasive, then histological subtype. General-purpose pathology foundation models learn transferable slide-level representations, but they are trained for breadth across organs rather than for this particular diagnostic hierarchy.
CorePath is a breast-specialized multimodal pathology foundation model built for that setting. Pathologists and computer scientists at West China Hospital of Sichuan University, with collaborators at Sichuan University, HKUST (Guangzhou), and three partner hospitals, adapted PRISM — the slide-level vision-language model of Shaikovski et al., arXiv:2405.10254, and not the later PRISM2 — using 7,901 paired breast CNB whole-slide images and diagnostic reports from two centers. The adaptation is deliberately narrow: only adapters inside the Perceiver slide encoder are trained, so breast specialization accrues in the visual pathway while the language embedding space inherited from pretraining is left intact.
The release also includes CorePath-CRG, a conformalized report-generation wrapper that decides, per case, whether a generated report is trustworthy enough to release at all. Both were described in an August 2026 arXiv preprint that has not yet been peer reviewed.
CorePath inherits PRISM's three-stage design: Virchow encodes tiles, a Perceiver slide encoder aggregates them into a slide embedding, and a BioGPT language branch handles text embedding, contrastive WSI-report alignment, and report decoding. Fine-tuning inserts bottleneck adapters (dimension 8, ReLU, residual scale 1.0) into feed-forward sublayers of the Perceiver — the FFN of the final transformer layer in the first block and the cross-attention FFNs of the first two blocks — leaving every other PRISM parameter frozen. Training ran 6 epochs on 8 NVIDIA RTX 3090 GPUs at an effective batch size of 640, learning rate 2e-4, with up to 2,560 tiles per slide and 768 text tokens. Chinese source reports were structured with DeepSeek-R1-Distill-Qwen-32B and translated to English with OpenBioLLM-70B before alignment.
On six independent private CNB cohorts, CorePath took the highest zero-shot weighted AUC in every cohort ahead of both PRISM and TITAN: 0.9669 to 0.9989 for cancer detection and 0.9643 to 0.9881 for invasion assessment, plus 0.9526 to 0.9735 for five-class histological subtyping in the five cohorts where subtype labels were available. On public benchmarks CorePath reached accuracy 0.8989 and weighted AUC 0.7780 for BCNB invasive carcinoma subtyping, and weighted AUCs of 0.8178 and 0.8252 for BRACS lesion stratification and fine-grained classification. In report generation, non-breast hallucinations fell from 30.1% to 2.8%; CorePath-CRG recorded none among released outputs, at the cost of full rejection rates between 19.15% and 38.58%.
The intended user is a pathology service drafting breast CNB reports at volume. CorePath supplies a slide embedding queryable with clinically defined categories, which suits workflows whose label spaces shift between screening, invasion assessment, and subtyping without a labeled training set for each. CorePath-CRG targets the deployment question that usually blocks generative reporting in clinical settings: it releases free text only when calibrated evidence supports it, degrades to a bare subtype label when the narrative is unreliable, and defers the case otherwise.
CorePath is evidence that a compact adapter fine-tune on a few thousand report-paired slides can specialize a general slide-level foundation model to one organ and specimen type while preserving zero-shot flexibility — a cheaper path than pretraining a new model per subspecialty. Pairing that with distribution-free risk control is the less common contribution: conformal methods have been applied to pathology classification and segmentation, but not to open-ended report generation. The study is retrospective, the fine-tuning data come from only two centers, and the exchangeability assumptions behind the conformal and Learn-Then-Test guarantees mean any new site needs local calibration. Code and weights are pledged to a public repository on acceptance; the repository is Apache-2.0 licensed but currently holds only a README, and the private WSI-report cohorts are not shareable.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.