Pith. sign in

REVIEW 3 major objections 4 minor 33 references

BRIGHT, a breast-specific pathology foundation model, outperforms generalist foundation models on 21 of 24 internal clinical tasks and 5 of 10 external tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 19:10 UTC pith:KEQ7AW32

load-bearing objection Large breast-pathology benchmark with a plausible but unverified central claim; the same-institution pretraining/test overlap has to be resolved before the 21/24 internal wins mean anything. the 3 major comments →

arxiv 2603.03030 v2 pith:KEQ7AW32 submitted 2026-03-03 cs.CV

BRIGHT: A Collaborative Generalist-Specialist Foundation Model for Breast Pathology

Xiaojing Guo , Jiatai Lin , Yumian Jia , Jingqi Huang , Zeyan Xu , Weidong Li , Longfei Wang , Jingjing Chen
show 34 more authors
Qin Li Weiwei Wang Lifang Cui Wen Yue Zhiqiang Cheng Xiaolong Wei Jianzhong Yu Xia Jin Baizhou Li Honghong Shen Jing Li Chunlan Li Yanfen Cui Yi Dai Yiling Yang Xiaolong Qian Liu Yang Yang Yang Guangshen Gao Yaqing Li Lili Zhai Chenying Liu Tianhua Zhang Zhenwei Shi Cheng Lu Xingchen Zhou Jing Xu Miaoqing Zhao Fang Mei Jiaojiao Zhou Ning Mao Fangfang Liu Chu Han Zaiyi Liu
This is my paper
classification cs.CV
keywords computational pathologyfoundation modelbreast pathologywhole-slide imagesgeneralist-specialistbiomarker predictionneoadjuvant therapy responsesurvival prediction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

BRIGHT is a foundation model built specifically for breast pathology, trained on more than 51,000 whole-slide images from over 40,000 patients. The paper's central claim is that a breast-specific model, formed by keeping a broad multi-organ generalist encoder frozen and adding a breast-adapted specialist encoder, generally beats leading generalist pathology models across the full clinical workflow: diagnosis, biomarker prediction, treatment response, and survival. To back this claim, the authors assembled the largest multi-institutional breast-pathology benchmark to date, with 24 internal and 10 external tasks. A sympathetic reader would care because it addresses two open questions at once: whether organ-specific depth is needed on top of generalist breadth, and whether an organ-specific foundation model can be built efficiently without starting from scratch.

Core claim

BRIGHT is the first pathology foundation model aimed at a single organ system, breast. Its architecture concatenates the frozen feature embeddings of a generalist encoder (trained on millions of pan-cancer slides) with the embeddings of a specialist encoder adapted to breast histology through low-rank fine-tuning of a subset of attention parameters. Trained on roughly 210 million tiles from 51,836 breast whole-slide images across 19 hospitals, BRIGHT is evaluated with a weakly supervised multiple-instance learning head on 24 clinical tasks spanning cancer detection, multi-class histological diagnosis, Nottingham grading, TIL assessment, node status, pTNM staging, eight biomarker and molecula

What carries the argument

The load-bearing mechanism is the dual-encoder collaboration: a frozen generalist encoder preserves universal morphological knowledge learned from large multi-organ data, while a parallel specialist encoder, fine-tuned with low-rank adapters on breast-only data, learns organ-specific patterns. The two 2560-dimensional feature vectors are concatenated into a single 5120-dimensional embedding used by a weakly supervised aggregator for each downstream task. This design is what lets BRIGHT claim both breadth and depth, and it is the component the paper identifies as a reusable template for organ-specific foundation models.

Load-bearing premise

The claim rests on the statement that every slide used for downstream validation was held out from BRIGHT's pretraining; because the pretraining data and the internal evaluation both come largely from the same hospital, any unrecognized patient or slide overlap would let the model score well by memorization, and the paper does not report a patient-ID overlap audit.

What would settle it

Compute the patient-ID intersection between the 42,876 pretraining slides from the primary institution and each downstream development/internal-validation set; if any patient appears in both, the internal superiority over generalist models could be explained by leakage. A cleaner test is to repeat the 24-task evaluation using only slides from hospitals outside the nineteen that contributed pretraining data and compare the performance gap to the internal results.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • BRIGHT reports AUCs of 0.992 for breast cancer detection and 0.973 for 10-class histological diagnosis on its internal natural-distribution cohort, suggesting the model could support screening triage.
  • On biomarker prediction, the paper estimates that at conservative 0.98 NPV/PPV thresholds, BRIGHT could eliminate the need for roughly 61% of ER, 92% of AR, and 47% of PD-L1 IHC assays, implying large cost and time savings if replicated prospectively.
  • BRIGHT predicts pathologic complete response to neoadjuvant therapy with AUC 0.786 internally and 0.641 externally, and it maintains performance in treatment-defined subgroups such as TNBC and HR-positive disease.
  • In survival analysis, BRIGHT-defined risk groups remain independent in multivariable Cox models (hazard ratios about 2.7–4.1), meaning the embeddings carry prognostic information beyond standard clinical variables.
  • The collaborative generalist-specialist architecture outperforms the specialist branch alone on most internal tasks, indicating that keeping the generalist pathway is beneficial rather than redundant.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: the same frozen-generalist-plus-adapted-specialist recipe can be tested on other organs; if it transfers, it would give a general route to organ-specific foundation models without expensive full pretraining.
  • Inference: the external results (5 of 10 top-1) are the weakest point; measuring performance on cohorts from hospitals completely absent from pretraining would separate genuine specialist generalization from adaptation to the primary institution's slide preparation.
  • Inference: the IHC-waiver proportions assume a pre-specified clinical threshold and a stable relationship between H&E appearance and biomarker status across sites; a prospective deployment study that tracks actual IHC ordering and error rates would test the practical benefit.
  • Inference: because BRIGHT concatenates features rather than distilling the specialist into the generalist, the final model is twice the embedding size of each branch; a follow-up could test whether a distilled single-encoder model retains the gains at lower computational cost.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces BRIGHT, a breast-specific pathology foundation model pretrained on roughly 51,000 breast WSIs from about 40,000 patients across 19 centers, using a dual-pathway design that concatenates frozen Virchow2 generalist embeddings with LoRA-adapted specialist embeddings. The authors evaluate the resulting features with a CLAM MIL aggregator across 24 clinical tasks spanning diagnosis, biomarkers, treatment response, and survival. The full text reports top-1 performance in 21 of 24 internal tasks and 5 of 10 external tasks, and claims that BRIGHT outperforms three generalist PFMs and its specialist-only variant. The abstract, however, reports different numbers: 25 of 25 internal tasks, 4 of 11 external tasks, and five comparison generalist PFMs. The central scientific claim is that the collaborative generalist-specialist fusion provides a scalable template for organ-specific foundation models, with the internal same-institution benchmark serving as the primary evidence.

Significance. If the holdout claim is verified, the study would be a substantial addition to computational pathology: it curates a large multi-institutional breast benchmark, demonstrates a parameter-efficient route from a generalist PFM to an organ-specific model, and evaluates an unusually broad clinical task spectrum with confidence intervals and public datasets (TCGA-BRCA, BRACS). The collaborative fusion idea is simple, reproducible in principle, and could transfer to other organ systems. At the same time, the external evidence is considerably weaker than the internal evidence, so the same-institution internal benchmark is load-bearing. The paper does not yet ship code or overlap-audit scripts, and the internal consistency problems must be resolved before the headline claims can be accepted.

major comments (3)
  1. [Methods — Benchmarking datasets; Results — internal validation] The claim that all evaluation WSIs are 'strictly held out from the BRIGHT foundation model’s pre-training phase' is the load-bearing assumption for the internal comparison. BRIGHT was pretrained on 42,876 WSIs from 32,054 patients at C1.TMUCIH, and the internal validation cohort is all consecutive 2024 breast biopsies at the same institution. The paper provides no patient-ID overlap audit, no slide-hash de-duplication record, and no statement of the pretraining collection dates. If any 2024 patients or their earlier slides appeared in pretraining, the reported 21/24 internal superiority over generalist PFMs could reflect memorization of the evaluation distribution, especially given that external top-1 performance is only 5/10. Please provide a formal patient-level (and, where available, slide-level) overlap audit for every internal development and validation cohort, and rerun the affecte
  2. [Abstract vs. Full Text (Introduction, Results, Methods)] The numerical claims are inconsistent: the abstract reports SOTA in 25 of 25 internal tasks and 4 of 11 external tasks, while the full text reports 21 of 24 internal and 5 of 10 external; the abstract says BRIGHT outperforms five leading generalist PFMs, while the Methods and Results compare three (Virchow2, UNIv2, CONCHv1.5) plus BRIGHT(S). These are not cosmetic discrepancies: the headline SOTA claim depends on the task counts, the number of baselines, and the numerator/denominator. Please reconcile the abstract, main text, figures, and supplementary task list, and report per-task results so readers can verify the counts.
  3. [Methods — Benchmarking datasets; Statistical analysis] The evaluation protocol is described in mutually incompatible ways. The Methods first define a fixed temporal holdout — all consecutive patients undergoing breast biopsy in calendar year 2024 at C1.TMUCIH — as the internal validation cohort. The next paragraph says that for each internal task the cohort was randomly split 80/20 at the patient level into training and validation sets. The Statistical analysis section then states that all performance metrics were based on 10-fold cross-validation. The paper needs to clarify which procedure generated the reported internal test results. If the 2024 cohort is the fixed test set, what is the random 80/20 split applied to? If results are from cross-validation, how is the 'natural, unselected distribution' and temporal holdout claim supported? Please provide the exact split/fold protocol and corresponding fold-level confidence intervals.
minor comments (4)
  1. [Figure 5 caption] The caption reads 'C26.TCGA-BRCA' in three places; this appears to be a typo for 'TCGA-BRCA'.
  2. [Results — biomarker prediction] The PD-L1 result (AUC 0.938 with 95% CI 0.814–1.000, trained on only 289 samples) has an extremely wide confidence interval. The statement that this 'underscores the model’s data-efficient nature' overstates what one task with a small sample can support.
  3. [Discussion] The phrase 'provides a delightful direction' is informal for a journal article; consider replacing with 'provides a promising direction' or similar.
  4. [Methods — Benchmarking datasets] The reported external hospital counts vary across sections: 4 external hospitals for diagnostic tasks, 6 for biomarker and NAT tasks, and 9 external hospitals overall. Please harmonize these numbers in the main text and supplementary tables.

Circularity Check

0 steps flagged

No significant circularity: BRIGHT's specialist features are learned by SSL on unlabeled tiles and the evaluation is an empirical comparison with external anchors; the same-institution overlap concern is a validity risk, not a demonstrated definitional reduction.

full rationale

The derivation chain is not circular. BRIGHT is built by LoRA fine-tuning a frozen generalist encoder (Virchow2) with DINOv2 on unlabeled breast WSIs, then concatenating generalist and specialist embeddings for downstream CLAM-based weakly supervised tasks. No downstream task label is used to construct the foundation-model features, and the claimed performance gains are empirical outcomes, not quantities fitted by construction. The manuscript states that 'all WSIs in both internal validation and external validation datasets used for downstream task evaluation are strictly held out from the BRIGHT foundation model’s pre-training phase,' and the evaluation is not self-referential: it includes external hospital cohorts, TCGA-BRCA, and BRACS, so the central claim does not reduce to the pretraining input. The main concern—that the internal validation cohort comes from the same institution that contributed 42,876 pretraining WSIs and that no patient-ID overlap audit is shown—is a legitimate data-hygiene/leakage risk, but it is not a demonstrated circular step. Under the instruction to claim circularity only with specific evidence of equation-level reduction or fitted-input-as-prediction, this risk alone does not raise the circularity score. No self-citation is load-bearing, no uniqueness theorem is imported from the authors, and no ansatz is smuggled in via citation.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The central empirical claims rest on domain assumptions about H&E signal, feature complementarity, label reliability, and the fairness of CLAM as the evaluation probe. The most fragile assumption is the unverified holdout of evaluation WSIs from pretraining. No new physical or mechanistic entities are introduced; the 'specialist pathway' is a LoRA adapter, not a new entity.

free parameters (2)
  • LoRA rank = 8
    Chosen architecture hyperparameter for the LoRA adapters; not justified by ablation and affects the specialist pathway's capacity and all downstream results.
  • NPV/PPV thresholds for IHC reduction estimate = 0.95-1.0 evaluated; headline threshold 0.98
    Thresholds applied post hoc to estimate the proportion of IHC assays that could be waived; the reported reduction rates are threshold-dependent and not part of model training.
axioms (5)
  • domain assumption H&E morphology encodes predictive signal for biomarkers, treatment response, and survival
    Central premise of all downstream tasks; supported by prior literature but not proven by this paper alone (Introduction, Results).
  • domain assumption Virchow2 generalist embeddings and LoRA-adapted specialist embeddings are complementary, and concatenation improves task performance
    Core architectural claim; tested only by comparing BRIGHT vs BRIGHT(S) vs Virchow2, not by ablating other fusion methods.
  • domain assumption The WSIs used for downstream evaluation were not included in BRIGHT pretraining
    Stated in Methods but not auditable with released artifacts; if false, internal SOTA results would be inflated.
  • domain assumption Pathology report labels are reliable ground truth
    All diagnostic and biomarker labels come from standardized pathology reports; no central pathology review is described.
  • domain assumption CLAM MIL aggregator with frozen encoders is a fair protocol for comparing PFMs
    A single aggregator is used; different PFMs could interact differently with CLAM, affecting the ranking.

pith-pipeline@v1.3.0-alltime-deepseek · 13137 in / 10649 out tokens · 96437 ms · 2026-08-02T19:10:18.918763+00:00 · methodology

0 comments
read the original abstract

Generalist pathology foundation models (PFMs), pretrained on large-scale multi-organ datasets, have demonstrated remarkable predictive capabilities across diverse clinical applications. However, their proficiency on the full spectrum of clinically essential tasks within a specific organ system remains an open question due to the lack of large-scale validation cohorts for a single organ as well as the absence of a tailored training paradigm that can effectively translate broad histomorphological knowledge into the organ-specific expertise required for specialist-level interpretation. In this study, we propose BRIGHT, the first PFM specifically designed for breast pathology, trained on over 51,000 breast whole-slide images derived from a cohort of over 40,000 patients across 19 hospitals. BRIGHT employs a collaborative generalist-specialist framework to capture both universal and organ-specific features. To comprehensively evaluate the performance of PFMs on breast oncology, we curate the largest multi-institutional cohorts to date for downstream task development and evaluation, comprising over 25,000 WSIs across 10 hospitals. The validation cohorts cover the full spectrum of breast pathology across 25 distinct clinical tasks spanning diagnosis, biomarker prediction, treatment response and survival prediction. Extensive experiments demonstrate that BRIGHT outperforms five leading generalist PFMs, achieving state-of-the-art (SOTA) performance in 25 of 25 internal validation tasks and in 4 of 11 external validation tasks with excellent heatmap interpretability. By evaluating on large-scale validation cohorts, this study not only demonstrates BRIGHT's clinical utility in breast oncology but also validates a collaborative generalist-specialist paradigm, providing a scalable template for developing PFMs on a specific organ system, accelerating the translation of foundation models into ...

Figures

Figures reproduced from arXiv: 2603.03030 by Baizhou Li, Cheng Lu, Chenying Liu, Chu Han, Chunlan Li, Fangfang Liu, Fang Mei, Guangshen Gao, Honghong Shen, Jianzhong Yu, Jiaojiao Zhou, Jiatai Lin, Jingjing Chen, Jing Li, Jingqi Huang, Jing Xu, Lifang Cui, Lili Zhai, Liu Yang, Longfei Wang, Miaoqing Zhao, Ning Mao, Qin Li, Tianhua Zhang, Weidong Li, Weiwei Wang, Wen Yue, Xia Jin, Xiaojing Guo, Xiaolong Qian, Xiaolong Wei, Xingchen Zhou, Yanfen Cui, Yang Yang, Yaqing Li, Yi Dai, Yiling Yang, Yumian Jia, Zaiyi Liu, Zeyan Xu, Zhenwei Shi, Zhiqiang Cheng.

Figure 1
Figure 1. Figure 1: Overview of BRIGHT study. a, Schematic of the clinical spectrum covered by the study, encompassing diagnostic subtyping, molecular biomarker prediction, treatment response assessment, and long-term survival prediction in breast oncology. b, Distribution of datasets used for the development and evaluation of each downstream clinical prediction task. BRIGHT covers wider clinical tasks and more benchmarking d… view at source ↗
Figure 2
Figure 2. Figure 2: Performance of PFMs on diagnostic tasks of breast pathology. a, Distribution of the downstream model development, internal test and external test datasets. The number in brackets is the number of hospitals for external validation. b, Model performance (AUROC) across individual diagnostic tasks on the internal validation set. The pie charts below indicate the total number of WSIs and the class distribution … view at source ↗
Figure 3
Figure 3. Figure 3: Performance of PFMs on molecular subtyping and biomarker prediction tasks of breast pathology. a, Distribution of the downstream model development, internal test and external test datasets. b, Model performance (AUROC) across individual tasks on the internal validation set. The values in brackets indicate the numbers of categories. c, Model performance (AUROC) on four biomarker prediction tasks and the mol… view at source ↗
Figure 4
Figure 4. Figure 4: Performance of PFMs in predicting breast cancer treatment response. a, Distribution of the downstream model development, internal test and external test datasets. b-c, Model performance (AUROC) across individual neoadjuvant therapy (NAT) response prediction tasks on the internal (b) and external validation sets (c). The values in brackets indicate the numbers of categories. d, Performance (AUROC) of the mo… view at source ↗
Figure 5
Figure 5. Figure 5: Performance of PFMs in prognostic prediction. a, Box plots comparing the C-index of BRIGHT with PFMs across three experimental settings: C26.TCGA-BRCA overall survival (OS), C26.TCGA-BRCA disease-free survival (DFS), and OS in the TNBC cohort of the C1.TMUCIH. Each box summarizes the C-index values obtained from all validation folds in 10-fold cross-validation. b, Kaplan-Meier survival curves for OS and DF… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 6 linked inside Pith

  1. [1]

    Annals Oncol.(2025)

    Marra, A.et al.Artificial intelligence entering the pathology arena in oncology: current applications and future perspectives. Annals Oncol.(2025). 2.LeCun, Y ., Bengio, Y . & Hinton, G. Deep learning.nature521, 436–444 (2015). 3.Lu, M. Y .et al.Ai-based pathology predicts origins for cancers of unknown primary.Nature594, 106–110 (2021)

    2025

  2. [4]

    medicine28, 154–163 (2022)

    Bulten, W.et al.Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge.Nat. medicine28, 154–163 (2022)

    2022

  3. [5]

    J.et al.Transformer-based biomarker prediction from colorectal cancer histology: A large-scale multicentric study.Cancer cell41, 1650–1661 (2023)

    Wagner, S. J.et al.Transformer-based biomarker prediction from colorectal cancer histology: A large-scale multicentric study.Cancer cell41, 1650–1661 (2023)

    2023

  4. [6]

    Medicine31, 3002–3010 (2025)

    Campanella, G.et al.Real-world deployment of a fine-tuned pathology foundation model for lung cancer biomarker detection.Nat. Medicine31, 3002–3010 (2025)

    2025

  5. [7]

    Heal.7(2025)

    Gustav, M.et al.Assessing genotype- phenotype correlations in colorectal cancer with deep learning: a multicentre cohort study.The Lancet Digit. Heal.7(2025)

    2025

  6. [8]

    medicine29, 135–146 (2023)

    Ogier du Terrail, J.et al.Federated learning for predicting histological response to neoadjuvant chemotherapy in triple-negative breast cancer.Nat. medicine29, 135–146 (2023)

    2023

  7. [9]

    Yuan, W.et al.Pancancer outcome prediction via a unified weakly supervised deep learning model.Signal transduction targeted therapy10, 285 (2025)

    2025

  8. [10]

    Skrede, O.-J.et al.Deep learning for prediction of colorectal cancer outcome: a discovery and validation study.The Lancet 395, 350–360 (2020)

    2020

  9. [11]

    medicine30, 85–97 (2024)

    Amgad, M.et al.A population-level digital histologic biomarker for enhanced prognosis of invasive breast cancer.Nat. medicine30, 85–97 (2024). 12.Oquab, M.et al.Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193(2023)

    Pith/arXiv arXiv 2024

  10. [13]

    InInternational conference on machine learning, 8748–8763 (PmLR, 2021)

    Radford, A.et al.Learning transferable visual models from natural language supervision. InInternational conference on machine learning, 8748–8763 (PmLR, 2021). 14.Moor, M.et al.Foundation models for generalist medical artificial intelligence.Nature616, 259–265 (2023)

    2021

  11. [15]

    Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T. J. & Zou, J. A visual–language foundation model for pathology image analysis using medical twitter.Nat. medicine29, 2307–2316 (2023)

    2023

  12. [16]

    17.Lu, M

    Wang, X.et al.A pathology foundation model for cancer diagnosis and prognosis prediction.Nature634, 970–978 (2024). 17.Lu, M. Y .et al.A visual-language foundation model for computational pathology.Nat. medicine30, 863–874 (2024). 18.Xiang, J.et al.A vision–language foundation model for precision oncology.Nature638, 769–778 (2025). 19.Lu, M. Y .et al.A mu...

    2024

  13. [21]

    J.et al.Towards a general-purpose foundation model for computational pathology.Nat

    Chen, R. J.et al.Towards a general-purpose foundation model for computational pathology.Nat. medicine30, 850–862 (2024)

    2024

  14. [22]

    medicine30, 2924–2935 (2024)

    V orontsov, E.et al.A foundation model for clinical-grade computational pathology and rare cancers detection.Nat. medicine30, 2924–2935 (2024)

    2024

  15. [23]

    Commun.16, 2366 (2025)

    Yang, Z.et al.A foundation model for generalizable cancer diagnosis and survival prediction from histopathological images.Nat. Commun.16, 2366 (2025)

    2025

  16. [24]

    Ma, J.et al.A generalizable pathology foundation model using a unified knowledge distillation pretraining framework. Nat. Biomed. Eng.1–20 (2025)

    2025

  17. [25]

    & Sung, J

    Xiong, C., Chen, H. & Sung, J. J. A survey of pathology foundation model: Progress and future directions.arXiv preprint arXiv:2504.04045(2025). 26.Mahmood, F. A benchmarking crisis in biomedical machine learning.Nat. Medicine31, 1060–1060 (2025)

    Pith/arXiv arXiv 2025

  18. [27]

    Ma, J.et al.Pathbench: A comprehensive comparison benchmark for pathology foundation models towards precision oncology.arXiv preprint arXiv:2505.20202(2025)

    Pith/arXiv arXiv 2025

  19. [28]

    Neidlinger, P.et al.Benchmarking foundation models as feature extractors for weakly supervised computational pathology. Nat. biomedical engineering1–11 (2025). 14/15

    2025

  20. [29]

    Commun.16, 3640 (2025)

    Campanella, G.et al.A clinical benchmark of public self-supervised pathology foundation models.Nat. Commun.16, 3640 (2025)

    2025

  21. [30]

    31.Hu, E

    Zimmermann, E.et al.Virchow2: Scaling self-supervised mixed magnification models in pathology.arXiv preprint arXiv:2408.00738(2024). 31.Hu, E. J.et al.Lora: Low-rank adaptation of large language models.ICLR1, 3 (2022)

    Pith/arXiv arXiv 2024

  22. [32]

    Y .et al.Data-efficient and weakly supervised computational pathology on whole-slide images.Nat

    Lu, M. Y .et al.Data-efficient and weakly supervised computational pathology on whole-slide images.Nat. biomedical engineering5, 555–570 (2021)

    2021

  23. [33]

    Brancati, N.et al.Bracs: A dataset for breast carcinoma subtyping in h&e histology images.Database2022, baac093 (2022)

    2022

  24. [34]

    Modi, S.et al.Trastuzumab deruxtecan in previously treated her2-low advanced breast cancer.New Engl. J. Medicine387, 9–20 (2022)

    2022

  25. [35]

    Wong, C.et al.Ai-driven mri biomarker for triple-class her2 expression classification in breast cancer: a large-scale multicenter study.Breast Cancer Res.27, 166 (2025)

    2025

  26. [36]

    Cortes, J.et al.Pembrolizumab plus chemotherapy in advanced triple-negative breast cancer.New Engl. J. Medicine387, 217–226 (2022)

    2022

  27. [37]

    35, 159–182 (2024)

    Loibl, S.et al.Early breast cancer: Esmo clinical practice guideline for diagnosis, treatment and follow-up.Annals Oncol. 35, 159–182 (2024)

    2024

  28. [38]

    Schmid, P.et al.Event-free survival with pembrolizumab in early triple-negative breast cancer.New Engl. J. Medicine386, 556–567 (2022)

    2022

  29. [39]

    40.Xu, Y .et al.A multimodal knowledge-enhanced whole-slide pathology foundation model.Nat

    El Bairi, K.et al.The tale of tils in breast cancer: a report from the international immuno-oncology biomarker working group.NPJ breast cancer7, 150 (2021). 40.Xu, Y .et al.A multimodal knowledge-enhanced whole-slide pathology foundation model.Nat. Commun.(2025). 41.Yan, S.et al.A multimodal vision foundation model for clinical dermatology.Nat. Medicine1–...

    2021

  30. [42]

    & Xie, P

    He, X., Zhang, Y ., Mou, L., Xing, E. & Xie, P. Pathvqa: 30000+ questions for medical visual question answering.arXiv preprint arXiv:2003.10286(2020)

    Pith/arXiv arXiv 2003

  31. [43]

    Li, Z.et al.Ai-enabled virtual spatial proteomics from histopathology for interpretable biomarker discovery in lung cancer. Nat. Medicine1–14 (2026)

    2026

  32. [44]

    Valanarasu, J. M. J.et al.Multimodal ai generates virtual population for tumor microenvironment modeling.Cell(2025)

    2025

  33. [45]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    Dosovitskiy, A. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020). 46.Weinstein, J. N.et al.The cancer genome atlas pan-cancer analysis project.Nat. genetics45, 1113–1120 (2013). 15/15

    Pith/arXiv arXiv 2010