Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Historical Report Guided Bi-modal Concurrent Learning for Pathology Report Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper argues that retrieving historical pathology reports and learning visual and textual tokens concurrently over weight-shared cross-attention yields state-of-the-art report generation and Her-2 prediction on PathText-BRCA.

desk verdict Useful architecture with a real new combination, but the SOTA margins rest on a split change and an unauditable Her-2 pipeline. read the letter →

arxiv 2506.18658 v1 pith:NQPGOQAK submitted 2025-06-23 cs.CV cs.AI

Ling Zhang , Boxiang Yun , Qingli Li , Yan Wang This is my paper · ORCID
classification cs.CVcs.AI
keywords pathologyreportgenerationwholeslideimageknowledgeretrievalcross-modallearningHer-2predictionvisual-languagepretrainingtransformermedicalcaptioning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes BiGen, a framework for generating pathology reports from whole-slide images (WSIs) that emulates how pathologists consult past cases. It builds a knowledge bank of sentence-level historical reports, retrieves sentences most similar to high-attention tissue regions of the current slide, and learns a compact visual token and a compact textual token together through weight-shared cross-attention. On the PathText-BRCA benchmark, BiGen reports state-of-the-art generation scores (BLEU-4 = 0.135, ROUGE-L = 0.293) and a Her-2 prediction F1 of 0.730, corresponding to a 7.4% relative improvement in NLP metrics and a 19.1% improvement in classification metrics over existing methods. If the results hold, historical reports can supply the semantic content missing from pixel features while the token design suppresses the redundancy of gigapixel slides.

What carries the argument

The load-bearing object is the pair of learnable tokens, the visual token $V$ and the textual token $T$, refined by cross-attention layers whose weights are shared between branches. A complementary mechanism is the knowledge retrieval: patch embeddings from PLIP are filtered to the top-$k$ regions by first-layer visual attention scores, spatially averaged, and matched by cosine similarity against a sentence-level knowledge bank built from training reports, returning the top-$v$ sentences per region.

What would settle it

A reader could take the PathText-BRCA test set, feed each slide with genuinely retrieved knowledge versus knowledge sampled at random from the knowledge bank, and check whether the reported gains (BLEU-4 0.135, Her-2 F1 0.730) collapse; if they do not, retrieval is not the source of the improvement.

Watch

Extended reading notes

Core claim

BiGen's central claim is that pathology report generation improves when the model is given access to historical diagnostic sentences retrieved by the visual evidence itself. The authors instantiate this with a bi-modal concurrent learning strategy: a learnable visual token iteratively cross-attends to patch features to distill key visual patterns, while a learnable textual token iteratively cross-attends to retrieved knowledge; the two branches share weights to align the modalities, and a multi-modal decoder generates the report from the concatenated tokens. On PathText-BRCA, this configuration beats seven existing captioning and report-generation methods on every reported metric, with the largest gains coming from weight sharing between the visual and knowledge branches.

Load-bearing premise

The retrieval step assumes that the first-layer visual cross-attention scores flag the diagnostically relevant patches, and that PLIP patch and sentence embeddings are aligned well enough that cosine similarity returns sentences carrying the same diagnostic content as the test slide.

Editorial extensions

If this is right

  • If the results are correct, BiGen outperforms the prior best pathology-report generator (HistGen) by 14.4% in BLEU-4 and improves Her-2 prediction F1 from 0.571 to 0.730 on PathText-BRCA.
  • Weight sharing between the visual and textual token cross-attention layers is what aligns the two modalities; removing it drops the average metric gain from 15.26% to 13.06%.
  • Knowledge retrieval without the textual token cross-attention hurts performance, showing that retrieved knowledge must be distilled to be useful.
  • Attention scores from the first visual cross-attention layer localize tumor tissue, so the same scores that guide retrieval also provide interpretable heatmaps.
  • The authors state that the paradigm can extend to other medical image interpretation tasks beyond pathology report generation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The attention-score-guided retrieval is effectively a weakly supervised selection of diagnostically relevant patches; this same signal could be reused for region-level interpretability in other whole-slide-image tasks.
  • On cohorts where PLIP's image-text alignment is weak or where historical reports are systematically biased, the retrieval step could inject noise rather than semantics; a testable safeguard would be to degrade or shuffle the knowledge bank and measure the drop.
  • A portion of the NLP gains may come from conditioning on text that is lexically similar to the target report rather than from better visual understanding; comparing against retrieval of random sentences would separate these effects.
  • The framework suggests a broader recipe: use a foundation model for retrieval and a small set of learnable tokens for cross-modal distillation, a pattern that may carry over to radiology or dermatology report generation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes BiGen, a framework for pathology report generation from whole slide images that combines a knowledge retrieval mechanism (retrieving sentence-level historical reports via PLIP embeddings of high-attention patches) with a bi-modal concurrent learning strategy using learnable visual and textual tokens and weight-shared cross-attention layers. Experiments on PathText-BRCA report state-of-the-art results in both NLP metrics (BLEU-4=0.135, ROUGE-L=0.293) and Her-2 classification metrics (F1=0.730), with a 7.4% relative improvement in NLP metrics and a 19.1% improvement in Her-2 F1 over prior methods. The authors also provide ablations over the main components and release code.

Significance. If the reported results are reproducible, BiGen is a useful contribution: it explicitly introduces historical-report retrieval into WSI report generation, proposes a relatively simple bi-modal token mechanism that improves over several strong baselines, and the public code and detailed ablations make the method easy to build on. The attention-heatmap visualization and the qualitative examples are also informative. However, the central empirical claims currently rest on a comparison protocol that is not clearly apples-to-apples and on an Her-2 evaluation pipeline that is not specified; these issues must be resolved before the significance of the reported margins can be assessed.

major comments (4)
  1. [§3.1, Dataset; Table 1] The paper states that it follows the PathText-BRCA splits of [3] and then 'Additionally, we remove duplicate patient samples across these sets, ensuring no patient overlap,' yielding 796/88/93 samples. It is not stated that any of the seven baselines in Table 1 were retrained or re-evaluated on this deduplicated split. If the baseline numbers come from the original splits, then the 7.4% NLP improvement and the 19.1% Her-2 F1 improvement are not measured on the same test set and the central SOTA claim is not supported. Please rerun all baselines on the exact deduplicated split with the same visual extractor, patch preprocessing, decoding strategy, and Her-2 extraction procedure, and report the resulting numbers.
  2. [§3.2, Table 1; Her-2 prediction metrics] The classification metrics Precision/Recall/F1 for Her-2 are computed from generated reports, but the manuscript never specifies how Her-2 status is extracted from free text (e.g., regex, keyword matching, or an external model) or whether the same extraction is applied to all baselines. Without this information, the Her-2 improvement cannot be audited. Moreover, there is a potential circularity: Fig. 2(e) shows that retrieved knowledge sentences can directly state Her-2 status (e.g., 'The tumor is ER-positive, PR-positive, and HER2-negative'), so the reported F1=0.730 may partly reflect copying biomarker labels from retrieved historical reports rather than genuinely predicting Her-2 from the WSI. Please specify the extraction pipeline, verify that it is identical for all compared methods, and provide an analysis of how much of the Her-2 gain depends on retrieved sentences that mention Her-2 status.
  3. [§4, Tables 1 and 2] All results in Tables 1 and 2 appear to come from a single run on a test set of 93 samples, with no error bars, standard deviations, or significance tests. Given the small test set and the moderate metric differences (e.g., BLEU-4 0.135 vs 0.118 for HistGen), the reported gains may be within run-to-run noise. Please report mean ± std over at least three random seeds and, where appropriate, bootstrap confidence intervals or a paired significance test.
  4. [§2.3, II; Fig. 3] The knowledge retrieval mechanism assumes that first-layer visual cross-attention scores identify diagnostically relevant patches and that PLIP patch embeddings and PLIP sentence embeddings are sufficiently aligned for cosine retrieval to return sentences with the same diagnostic content as the test WSI. This assumption is plausible but is only validated indirectly through final metric improvements on a single dataset. If the attention scores or embedding alignment degrade on a new cohort, retrieved knowledge could become noise, and the reported gains would not transfer. A more direct validation, such as an analysis of retrieval precision on held-out cases or a comparison of retrieved sentences against the ground-truth report content, would strengthen the paper.
minor comments (6)
  1. [§2.2 and §2.3, III] The visual token cross-attention is applied for L layers in Eq. (1), while the textual token cross-attention is applied for L−1 layers in Eq. (2), because the first-layer visual attention scores are needed for retrieval. Please clarify this asymmetry explicitly in the text, since a reader could otherwise expect both branches to use the same number of layers.
  2. [§2.3, II] The notation for the top-k selection is confusing: k is described as a ratio, but the set P is written as {p_i}_{i=1}^{M×k}, which suggests an integer count. Please define the number of selected patches as ⌊M·k⌋ and ensure all subsequent dimensions are consistent.
  3. [§3.2] There is a typo: 'genaration' should be 'generation' in the sentence introducing MI-Gen and HistGen.
  4. [§3.3] The phrase 'irrelative patches' should be 'irrelevant patches', and 'representatively' in the ablation discussion should likely be 'respectively'.
  5. [Fig. 1 and Fig. 2] Several figure labels are garbled or incomplete, for example the subscripted tokens in Fig. 1 and the caption 'Correctly prediced information' in Fig. 2. Please proofread the figures and captions.
  6. [Fig. 3] The sensitivity plots do not specify which metric is shown; please state that the y-axis is, for example, BLEU-4, and whether other metrics show the same trend.

Circularity Check

1 steps flagged · score 6.0 of 10

The Her-2 'prediction' headline (F1=0.730, +19.1%) partially reduces by construction: retrieved knowledge sentences from the training-report bank directly state Her-2 status, so the classification metric reflects retrieval of known labels as much as visual diagnosis; the NLP report-generation claim is independently evaluated and not circular.

  1. fitted input called prediction [Sec. 2.3 (Knowledge Bank Construction I, Knowledge Retrieval II), Sec. 3.2 (Table 1 Her-2 metrics), Fig. 2(e)]
    "...the original reports in the training set (e.g., PathText (BRCA)) are split into individual sentences and encoded using the text encoder of PLIP. The resulting sentence embeddings are stored as the knowledge bank... (Sec. 2.3-I). '...we adopt three classification metrics to evaluate Her-2 prediction in generated reports: Precision, Recall and F1 score' (Sec. 3.2). Fig. 2(e) retrieved knowledge item 5: 'The tumor is ER-positive, PR-positive, and HER2-negative.'"

    The headline Her-2 F1=0.730 is reported as classification of the generated report text, but the generated text is decoded from F=Concat(VL,T^{L-1}) (Sec. 2.4), where T^{L-1} is a textual token refined over the retrieved knowledge R (Eq. 2), and R is obtained by cosine retrieval (Sec. 2.3-II) from a knowledge bank of training-set report sentences that explicitly state the evaluation label (Fig. 2(e): 'HER2-negative', 'ER-positive, PR-positive'). The Her-2 answer is thus present in the model's input before any classification happens; the decoder can transcribe the retrieved sentence into the report, making the 'prediction' a retrieval of known training-set labels rather than a visual diagnosis.

full rationale

The central NLP claim (BLEU-4=0.135, ROUGE-L=0.293, Table 1) is not circular: BiGen is trained with an NLL objective against ground-truth reports and evaluated on a held-out test set (796/88/93 after patient deduplication, Sec. 3.1) with standard captioning metrics, so the generation numbers measure held-out fidelity. The framework contains no load-bearing self-citation: all cited components (MI-Gen's splits [3], CLAM [17], UNI [4], PLIP [11], HistGen [10], the early-attention inspiration [2]) are independent external works, and none of the authors' own prior results are invoked to justify a design choice or forbid an alternative. The one genuinely circular channel is the Her-2 headline: the knowledge bank is built from training-report sentences that directly state Her-2 status, retrieval (Sec. 2.3-II) returns those label-bearing sentences to the textual token (Eq. 2), the decoder conditions on them (Eq. 3), and Table 1's Precision/Recall/F1 are described as 'Her-2 prediction in generated reports.' The 19.1% Her-2 F1 gain over retrieval-free baselines therefore reflects, in part, transcription of retrieved training labels rather than visual diagnosis, and no separate visual Her-2 estimator or text-extraction pipeline is specified, making this metric partially circular by construction. Not counted as circularity (but noted as validity risks): the deduplicated split (Sec. 3.1) may differ from the splits on which baselines were scored, and hyperparameters k, v, m are tuned on the same benchmark (Fig. 3); these affect comparability and selection bias, not the derivation chain. Weighing the vulnerable Her-2 claim (one of two headline results) against the independently evaluated NLP claim, the overall circularity score is 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The core new machinery is the learnable visual token and textual token, which are trainable parameters rather than invented physical entities. The free hyperparameters k, m, v are hand-set and tuned on this dataset. The main assumptions are about PLIP embedding quality, attention-score patch selection, and spatial pooling.

free parameters (3)
  • patch selection ratio k = 0.4
    Chosen by sensitivity study in Section 3.3; setting k=1 (all patches) degrades performance, and k is stable in [0.4,0.8].
  • tissue region size m = 20
    Chosen by sensitivity study in Section 3.3; performance is stable for m in [20,80], with smaller m preferred for lower computation.
  • number of retrieved knowledge features v = 3
    Chosen by sensitivity study in Section 3.3; best performance is achieved as v changes from 3 to 7.
assumptions (4)
  • domain assumption PLIP provides a shared embedding space where pathology image patches and pathology sentences have comparable representations.
    Section 2.3 items I and II use PLIP text and image encoders to build the knowledge bank and to retrieve by cosine similarity; if these embeddings are not semantically aligned, retrieval is meaningless.
  • ad hoc to paper First-layer visual cross-attention attention scores mark the diagnostically relevant patches.
    Section 2.3 item II selects top-k patches by attention scores from the first VTCA layer to query the knowledge bank; this selection is not independently validated.
  • domain assumption Spatially neighboring patches on a WSI tend to share tissue type, so averaging within regions preserves diagnostic content.
    Section 2.3 item II uses uniform region pooling of size m to compress patch embeddings before retrieval; if regions mix different tissue types, diagnostic information may be lost.
  • domain assumption Sentence-level splitting of reports preserves enough context for knowledge retrieval.
    Section 2.3 item I builds the knowledge bank from individual sentences due to PLIP token limits; if long-range report context is needed, sentence-level retrieval may be insufficient.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Historical Report Guided Bi-modal Concurrent Learning for Pathology Report Generation." pith.science (2026). https://pith.science/paper/NQPGOQAK

@misc{pith2026250618658,
  author       = {Pith},
  title        = {Pith review of: Historical Report Guided Bi-modal Concurrent Learning for Pathology Report Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NQPGOQAK}},
  note         = {Machine review of arXiv:2506.18658}
}
read the original abstract

Automated pathology report generation from Whole Slide Images (WSIs) faces two key challenges: (1) lack of semantic content in visual features and (2) inherent information redundancy in WSIs. To address these issues, we propose a novel Historical Report Guided \textbf{Bi}-modal Concurrent Learning Framework for Pathology Report \textbf{Gen}eration (BiGen) emulating pathologists' diagnostic reasoning, consisting of: (1) A knowledge retrieval mechanism to provide rich semantic content, which retrieves WSI-relevant knowledge from pre-built medical knowledge bank by matching high-attention patches and (2) A bi-modal concurrent learning strategy instantiated via a learnable visual token and a learnable textual token to dynamically extract key visual features and retrieved knowledge, where weight-shared layers enable cross-modal alignment between visual features and knowledge features. Our multi-modal decoder integrates both modals for comprehensive diagnostic reports generation. Experiments on the PathText (BRCA) dataset demonstrate our framework's superiority, achieving state-of-the-art performance with 7.4\% relative improvement in NLP metrics and 19.1\% enhancement in classification metrics for Her-2 prediction versus existing methods. Ablation studies validate the necessity of our proposed modules, highlighting our method's ability to provide WSI-relevant rich semantic content and suppress information redundancy in WSIs. Code is publicly available at https://github.com/DeepMed-Lab-ECNU/BiGen.

Figures

Figures reproduced from arXiv: 2506.18658 by the authors.

Figure 1
Figure 1. Overview of BiGen. The overall framework in (a), consists of a bi-modal concur￾rent learning encoder and a multi-modal decoder. "CA" and "FFN" indicates Cross￾attention and Feed-Forward Network. "VTCA" and "TTCA" refer to the visual and textual token cross-attention. We construct a knowledge bank in (b) and retrieve WSI￾relevant knowledge from the knowledge bank through high-attention patches in (c). 2 Method Our go… view at source ↗
Figure 2
Figure 2. Visualization of the (a) thumbnail, (b) attention score heatmap, (c) report ground-truth, (d) prediction and (e) retrieved knowledge of sample "TCGA-A8-A08L". generated words {y l i }i<n serve as queries, the fused representation F provides keys and values for the decoder and the n-th word y l n is predicted as: y l n = CrossAttn({y l i}i<n, F; Ωl), l ∈ {1, 2, . . . , L}, (3) where Ωl indicates learnable weights for… view at source ↗
Figure 3
Figure 3. Performance changes by varying the (a) selecting ratio k, (b) number of knowl￾edge features v and region size m. 3.3 Ablation Study We conduct ablation studies on five components: 1)Weight Sharing (WS) be￾tween the visual and knowledge branch, 2) Weight Sharing across Layers (WSL) in each branch, 3) Visual Token Cross-attention (VTCA), 4) Knowledge Re￾trieval (KR) and 5) Textual Token Cross-attention (TTCA). As show… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages

  1. [3]

    In: MIC- CAI (2024)

    Chen, P., Li, H., Zhu, C., Zheng, S., Shui, Z., Yang, L.: Wsicaption: Multiple instance generation of pathology reports for gigapixel whole-slide images. In: MIC- CAI (2024)

    work page 2024

  2. [1]

    In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summa- rization (2005)

    Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with im- proved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summa- rization (2005)

    work page 2005

  3. [2]

    In: ECCV (2024)

    Chen, L., Zhao, H., Liu, T., Bai, S., Lin, J., Zhou, C., Chang, B.: An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision- language models. In: ECCV (2024)

    work page 2024

  4. [4]

    Nature Medicine30(2024)

    Chen, R.J., Ding, T., Lu, M.Y., Williamson, D.F., Jaume, G., Song, A.H., Chen, B., Zhang, A., Shao, D., Shaban, M., et al.: Towards a general-purpose foundation model for computational pathology. Nature Medicine30(2024)

    work page 2024

  5. [5]

    arXiv preprint arXiv:2204.13258 (2022)

    Chen, Z., Shen, Y., Song, Y., Wan, X.: Cross-modal memory networks for radiology report generation. arXiv preprint arXiv:2204.13258 (2022)

    arXiv 2022

  6. [6]

    arXiv preprint arXiv:2010.16056 (2020)

    Chen, Z., Song, Y., Chang, T.H., Wan, X.: Generating radiology reports via memory-driven transformer. arXiv preprint arXiv:2010.16056 (2020)

    arXiv 2020

  7. [7]

    Nature Protocols20(2025)

    El Nahhas, O.S., van Treeck, M., Wölflein, G., Unger, M., Ligero, M., Lenz, T., Wagner, S.J., Hewitt, K.J., Khader, F., Foersch, S., et al.: From whole-slide image to biomarker prediction: end-to-end weakly supervised deep learning in computa- tional pathology. Nature Protocols20(2025)

    work page 2025

  8. [8]

    World health organization (2000)

    Fritz, A.G.: International classification of diseases for oncology: ICD-O. World health organization (2000)

    work page 2000

Show all 29 references
  1. [9]

    In: WACV (2024)

    Gu, T., Liu, D., Li, Z., Cai, W.: Complex organ mask guided radiology report generation. In: WACV (2024)

    2024

  2. [10]

    In: MICCAI (2024)

    Guo, Z., Ma, J., Xu, Y., Wang, Y., Wang, L., Chen, H.: Histgen: Histopathol- ogy report generation via local-global feature encoding and cross-modal context interaction. In: MICCAI (2024)

    2024

  3. [11]

    Nature medicine29(2023) 10 Ling Zhang, Boxiang Yun, Qingli Li, and Yan Wang ()

    Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual–language foundation model for pathology image analysis using medical twitter. Nature medicine29(2023) 10 Ling Zhang, Boxiang Yun, Qingli Li, and Yan Wang ()

    2023

  4. [12]

    NIPS36(2024)

    Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. NIPS36(2024)

    2024

  5. [13]

    In: CVPR (2023)

    Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task- specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: CVPR (2023)

    2023

  6. [14]

    In: MICCAI (2021)

    Liu, Q., Louis, P.C., Lu, Y., Jha, A., Zhao, M., Deng, R., Yao, T., Roland, J.T., Yang, H., Zhao, S., et al.: Simtriplet: Simple triplet representation learning with a single gpu. In: MICCAI (2021)

    2021

  7. [15]

    In: CVPR (2017)

    Lu,J.,Xiong,C.,Parikh,D.,Socher,R.:Knowingwhentolook:Adaptiveattention via a visual sentinel for image captioning. In: CVPR (2017)

    2017

  8. [16]

    Nature634(2024)

    Lu, M.Y., Chen, B., Williamson, D.F., Chen, R.J., Zhao, M., Chow, A.K., Ikemura, K., Kim, A., Pouli, D., Patel, A., et al.: A multimodal generative ai copilot for human pathology. Nature634(2024)

    2024

  9. [17]

    Nature biomedical engineering5(2021)

    Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering5(2021)

    2021

  10. [18]

    arXiv preprint arXiv:2010.10042 (2020)

    Miura, Y., Zhang, Y., Tsai, E.B., Langlotz, C.P., Jurafsky, D.: Improving factual completeness and consistency of image-to-text radiology report generation. arXiv preprint arXiv:2010.10042 (2020)

    2020 arXiv

  11. [19]

    In: ACL (2002)

    Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: ACL (2002)

    2002

  12. [20]

    In: Proceedings of Workshop on Text Summarization of ACL, Spain (2004)

    Rouge, L.C.: A package for automatic evaluation of summaries. In: Proceedings of Workshop on Text Summarization of ACL, Spain (2004)

    2004

  13. [21]

    In: MICCAI (2024)

    Tan, J.W., Kim, S., Kim, E., Lee, S.H., Ahn, S., Jeong, W.K.: Clinical-grade multi- organ pathology report generation for multi-scale whole slide images via a seman- tically guided medical text foundation model. In: MICCAI (2024)

    2024

  14. [22]

    medRxiv (2024)

    Tran, M., Schmidle, P., Wagner, S.J., Koch, V., Lupperger, V., Feuchtinger, A., Boehner, A., Kaczmarczyk, R., Biedermann, T., Eyerich, K., et al.: Generating highly accurate pathology reports from gigapixel whole slide images with histogpt. medRxiv (2024)

    2024

  15. [23]

    NEJM AI1(2024)

    Tu, T., Azizi, S., Driess, D., Schaekermann, M., Amin, M., Chang, P.C., Carroll, A., Lau, C., Tanno, R., Ktena, I., et al.: Towards generalist biomedical ai. NEJM AI1(2024)

    2024

  16. [24]

    NIPS (2017)

    Vaswani, A.: Attention is all you need. NIPS (2017)

    2017

  17. [25]

    In: CVPR (2015)

    Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: A neural image caption generator. In: CVPR (2015)

    2015

  18. [26]

    Nature (2024)

    Xu, H., Usuyama, N., Bagga, J., Zhang, S., Rao, R., Naumann, T., Wong, C., Gero, Z., González, J., Gu, Y., et al.: A whole-slide foundation model for digital pathology from real-world data. Nature (2024)

    2024

  19. [27]

    In: ICML (2015)

    Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: ICML (2015)

    2015

  20. [28]

    In: CVPR (2022)

    Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y.: Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathol- ogy whole slide image classification. In: CVPR (2022)

    2022

  21. [29]

    In: CVPR (2024)

    Zhu, L., Wei, F., Lu, Y.: Beyond text: Frozen large language models in visual signal comprehension. In: CVPR (2024)

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.