Pith. sign in

REVIEW 4 major objections 5 minor 39 references

Single GPU Task Adaptation of Pathology Foundation Models for Whole Slide Image Analysis

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TAPFM adapts pathology foundation models to clinical slide tasks on a single GPU.

desk verdict The central loss in Eq. 5 has a sign error that, as written, makes the feature alignment term increase the classification loss; the paper needs a major revision, but the single-GPU adaptation idea and the evaluation are worth engaging with. read the letter →

arxiv 2506.05184 v1 pith:CXLECUS5 submitted 2025-06-05 cs.CV

Neeraj Kumar , Swaraj Nanda , Siddharth Singi , Jamal Benhamida , David Kim , Jie-Fu Chen , Amir Momeni-Boroujeni , Gregory M. Goldgof
show 2 more authors
Gabriele Campanella Chad Vanderbilt
This is my paper · ORCID
classification cs.CV
keywords pathologyfoundationmodelmultipleinstancelearningwholeslideimagemutationpredictionvisiontransformerattentionsingle-GPUfine-tuningcomputationaltaskadaptation
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

This paper proposes TAPFM, a way to adapt a pretrained pathology foundation model to a clinical whole-slide task on a single GPU. Instead of treating the foundation model as a fixed feature extractor or fine-tuning it end-to-end, TAPFM uses the vision transformer's own CLS attention to weight tile contributions in multiple instance learning, and splits optimization into two detached stages: one updates the MIL aggregator, the other updates the foundation model with a task adaptation loss. On mutation prediction for bladder cancer (FGFR3) and lung adenocarcinoma (EGFR, KRAS, MET, ALK), TAPFM consistently outperforms both fixed-feature MIL aggregators and fine-tuned baselines, with H-Optimus-0 reaching AUC 0.9021 on TCGA BLCA FGFR3. The message is that large pathology foundation models can be practically adapted for clinical tasks without multi-GPU resources.

What carries the argument

The bag representation is $Z = \sum_{i=1}^{K} a_i z_i$, where $z_i$ is the CLS feature of tile $i$ and $a_i$ is the average attention the CLS token pays to that tile's patches across heads and tokens, min-max scaled and softmax normalized. The engine is the task adaptation loss $L_{PFM} = L_{feature} + \lambda L_{attention}$, computed from gradients detached from the aggregator: $L_{feature} = -\operatorname{tr}(Z G_z^T)$ moves features along the direction that reduces the classification loss, and $L_{attention} = a^T g_a$ reinforces or suppresses tile weights based on their informativeness. Separate computational graphs for the aggregator and PFM, with detached features, attention, and gradients, are what make this fit on one GPU.

What would settle it

Take a WSI dataset with tile-level annotations of diagnostically relevant regions, run TAPFM on a mutation-prediction task, and check whether the learned attention weights $a_i$ assign higher mass to annotated tumor regions; if the weights do not correlate with diagnostic relevance while AUC still improves, the attention mechanism is not doing the claimed work. Alternatively, replace the attention weights with random weights while keeping the dual-loss updates; if performance does not drop, the aggregation mechanism is not the source of gains.

Watch

Extended reading notes

Core claim

The central claim is that task adaptation of a pathology foundation model can be driven by its own attention mechanism, and that decoupling the PFM and aggregator optimization stabilizes training. For each tile, the average CLS attention across heads and tokens defines a tile importance weight; the bag representation is the attention-weighted sum of tile features. The aggregator is trained with a weighted cross-entropy loss, while the PFM is updated with a task adaptation loss that combines a feature alignment term and an attention loss term. Experiments report that TAPFM outperforms fixed-feature MIL (ABMIL, CLAM, DSMIL, VarMIL) and end-to-end fine-tuned versions of the same models on both institutional and TCGA cohorts, and that H-Optimus-0 (TAPFM) achieves the best AUC in every comparison.

Load-bearing premise

The method assumes that the average attention a tile's CLS token gives to its own patches reflects how important that tile is for the slide-level task; if that correlation fails, the MIL weights are not meaningful and any improvement would come from the feature alignment loss alone.

Editorial extensions

If this is right

  • If the claim holds, any ViT-based pathology foundation model can be adapted to a new clinical task without a multi-GPU cluster, since training fits on a single H100 with 80 GB.
  • Mutation prediction from H&E slides could be run at scan time, giving molecularly informed results in hours rather than days, and the same model can predict several actionable mutations at once, including rare ones like MET and ALK.
  • The approach extends to other bag-level clinical endpoints beyond mutations, such as survival, recurrence, or treatment response, since the mechanism does not depend on the specific label.
  • Attention weights learned this way could be reused as tile-level supervision, for example to localize diagnostic regions or guide pathologist review.
  • The reported convergence within 4 to 11 epochs means adaptation is cheap enough to iterate on per-institution cohorts and scanners.

Reading between the lines

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

  • The paper does not isolate whether the gains come from attention-based aggregation, the feature alignment loss, or the detachment; an ablation that replaces the attention weights with random or learned weights while keeping the dual-loss structure would tell.
  • If CLS attention does not reliably mark diagnostic tiles on some tasks, TAPFM's aggregation could be biased; a test against tile-level annotations or expert-marked regions would reveal this.
  • The same dual-graph detachment idea could be applied to other weakly supervised aggregators, such as graph or transformer MIL, not just linear attention-weighted sums.
  • The externally validated TCGA results suggest the method transfers across scanners and institutions, but only for two tumor types; the strongest test would be prospective validation with multiple sites and varied slide preparation.
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 / 5 minor

Summary. The paper proposes TAPFM, a single-GPU task-adaptation method for pathology foundation models in whole slide image analysis. It uses the CLS attention of a ViT-based PFM as MIL aggregation weights, detaches tile features and attention weights from the PFM computation graph, and fine-tunes the PFM with a task adaptation loss composed of a feature alignment term and an attention term. The method is evaluated on FGFR3 mutation prediction in bladder cancer and EGFR/KRAS/MET/ALK prediction in lung adenocarcinoma, using both institutional and TCGA cohorts, and reports AUC improvements over fixed-feature MIL aggregators and end-to-end fine-tuned baselines.

Significance. If the reported results are reproducible, the practical contribution is genuine: the paper demonstrates that large PFMs can be adapted to clinically relevant mutation-prediction tasks on a single H100 GPU, with external TCGA validation and a public implementation. The ablations for the attention-loss weight and tile sampling density are useful, and the computational-complexity analysis is a helpful addition. However, the central optimization equation contains a sign inconsistency that currently prevents the reader from verifying the proposed mechanism, and the empirical claims lack uncertainty quantification. The theoretical support in Section 3.4 is asserted rather than rigorously proven. These issues need to be addressed before the contribution can be fully assessed.

major comments (4)
  1. [Section 3.3, Eq. (5)] There is a sign error in the definition of L_feature. The left-hand side, -tr(Z G_z^T), equals -Σ_i ⟨z_i, g_z_i⟩ under the standard inner-product convention, not +Σ_i ⟨z_i, g_z_i⟩ as written. Since ∇_{z_i}(-Σ_j ⟨z_j, g_z_j⟩) = -g_z_i, minimizing this loss moves z_i in the direction +η g_z_i, which increases L_agg to first order, the opposite of the stated goal that the loss 'guides feature vectors to move in the direction that reduces the classification loss.' If the code actually uses +tr(Z G_z^T), then Eq. (5) is a typo and the mechanism is correct; if the code uses -tr, the reported gains cannot be attributed to Eq. (7). Please correct the sign, verify which expression is implemented, and include a one-line first-order derivation showing which direction minimizes L_agg.
  2. [Section 3.2, Eqs. (1)-(2)] The attention weights used for MIL aggregation are averaged CLS attention of each tile to its own patches, not a learned bag-level importance. It is not self-evident that these intra-tile values, after min-max scaling and softmax normalization, should serve as tile weights for the slide-level task. This is a load-bearing assumption because the bag representation in Eq. (2) and the attention loss in Eq. (6) both depend on it. Please provide an ablation or diagnostic that supports the semantic link between intra-tile attention and tile-level diagnostic relevance, for example by replacing a with uniform weights or with ABMIL-style learned attention while keeping the same dual-loss PFM updates.
  3. [Section 4.1, Table 1] All comparisons are reported as point estimates of AUC from a single test split. Several differences between the top methods are small (e.g., 0.9021 vs. 0.8994 for TCGA BLCA; 0.8553 vs. 0.8528 for TCGA LUAD EGFR). Without multiple seeds, bootstrap confidence intervals, or a paired significance test, the claim that TAPFM 'consistently outperforms' the baselines is not statistically supported. Please add uncertainty estimates, or explicitly state that the observed differences are descriptive and not tested for significance.
  4. [Section 3.4, Proposition 1] The proof of Proposition 1 restates the updating rules in Eqs. (12)-(15) and asserts that the parameter trajectories are more stable, but it does not define a notion of trajectory stability or provide a mathematical argument that the detached updates are more stable than joint optimization. As written, the proposition is not established. Either replace it with a precise theorem that includes conditions on the losses and learning rates, or present this section as a heuristic motivation rather than a formal proof.
minor comments (5)
  1. [Section 4] The acronym 'TAPM' appears several times (e.g., 'the proposed TAPM approach') where 'TAPFM' is intended; please correct these instances.
  2. [Section 4, Implementation Details] There is a typo in 'AdamW ... is used as opitimzer'; it should be 'optimizer'.
  3. [Figure 3 caption] The word 'bianry' should be 'binary'.
  4. [Appendix B] The references to 'Appendix 4a' and 'Appendix 4b' should refer to 'Figure 4a' and 'Figure 4b', respectively.
  5. [Section 3.2, Eq. (2)] The symbol Z is used both for the K×D feature matrix and for the bag-level representation after aggregation. Using a distinct notation such as z_bag would avoid confusion.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity; empirical evaluation is external and self-contained.

full rationale

The paper's core outcome—TAPFM outperforming fixed-feature and fine-tuned baselines on BLCA/LUAD mutation prediction—is established by held-out institutional splits and external TCGA cohorts, with no fitted parameter being reused as a 'prediction' of the same quantity. The task-adaptation losses in Eqs. 5-7 are optimization objectives constructed from detached aggregator gradients (first-order surrogates), not self-referential predictions; they do not encode the test label as an input. Self-citations such as DEMO [31] and the authors' benchmark [13] are used only as background motivation for single-GPU adaptation and are not load-bearing for the reported AUCs. The proof of Proposition 1 largely restates the detach-based update rules rather than independently establishing stability, and the apparent sign inconsistency in Eq. 5 is a mathematical correctness concern, but neither is a circular reduction of the paper's central empirical claim.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that intra-tile attention weights are meaningful for MIL aggregation, that the detached-gradient surrogate losses are valid optimization directions, and that TCGA-pretrained PFMs do not compromise the external TCGA validation. No new physical or algorithmic entities are introduced beyond the training procedure itself.

free parameters (2)
  • lambda (attention loss weight) = 1.0
    Equation 7; chosen via ablation on the BLCA FGFR3 validation set (Section 4.4).
  • tiles per WSI per epoch for each PFM = 300 (UNI), 100 (GigaPath), 75 (H-Optimus-0)
    Maximum processable on a single H100 GPU; not tuned for performance but affects results.
assumptions (4)
  • domain assumption Averaged CLS attention across patches within a tile is a meaningful proxy for that tile's importance to the bag-level prediction.
    Equation 1 is used to define bag weights; no evidence is given that intra-tile attention correlates with diagnostic relevance.
  • domain assumption The first-order approximation of the aggregator loss via detached gradients provides a valid update direction for the PFM.
    Equations 5 and 6 define linear surrogates; their validity for multi-step optimization is not established.
  • domain assumption Detaching gradients between aggregator and PFM produces stable training and prevents catastrophic forgetting.
    Proposition 1 and Appendix A.4 make this claim without rigorous proof or empirical demonstration.
  • domain assumption Pretraining of PFMs on TCGA does not invalidate external validation on TCGA cohorts.
    UNI, GigaPath, and H-Optimus-0 may have been pretrained on TCGA slides; the paper does not discuss this potential leakage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Single GPU Task Adaptation of Pathology Foundation Models for Whole Slide Image Analysis." pith.science (2026). https://pith.science/paper/CXLECUS5

@misc{pith2026250605184,
  author       = {Pith},
  title        = {Pith review of: Single GPU Task Adaptation of Pathology Foundation Models for Whole Slide Image Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CXLECUS5}},
  note         = {Machine review of arXiv:2506.05184}
}
read the original abstract

Pathology foundation models (PFMs) have emerged as powerful tools for analyzing whole slide images (WSIs). However, adapting these pretrained PFMs for specific clinical tasks presents considerable challenges, primarily due to the availability of only weak (WSI-level) labels for gigapixel images, necessitating multiple instance learning (MIL) paradigm for effective WSI analysis. This paper proposes a novel approach for single-GPU \textbf{T}ask \textbf{A}daptation of \textbf{PFM}s (TAPFM) that uses vision transformer (\vit) attention for MIL aggregation while optimizing both for feature representations and attention weights. The proposed approach maintains separate computational graphs for MIL aggregator and the PFM to create stable training dynamics that align with downstream task objectives during end-to-end adaptation. Evaluated on mutation prediction tasks for bladder cancer and lung adenocarcinoma across institutional and TCGA cohorts, TAPFM consistently outperforms conventional approaches, with H-Optimus-0 (TAPFM) outperforming the benchmarks. TAPFM effectively handles multi-label classification of actionable mutations as well. Thus, TAPFM makes adaptation of powerful pre-trained PFMs practical on standard hardware for various clinical applications.

Figures

Figures reproduced from arXiv: 2506.05184 by the authors.

Figure 1
Figure 1. WSI Processing Pipeline – A representative H&E stained WSI scanned at 40× optical resolution with 100, 000 × 125, 000 RGB pixels. For efficient processing, manageable sized tiles (e.g. 224 × 224 × 3) shown in black boxes are usually extracted from the tissue region. Bag of tiles obtained from a WSI are then passed through a pathology foundation model to obtain feature representations (one feature vector per tile) wh… view at source ↗
Figure 2
Figure 2. TAPFM Algorithm Overview – The proposed approach processes WSI tiles through a pathology foundation model (PFM) to extract features Z and attention weights a (solid arrows). These are detached from the PFM’s computation graph and used by the MIL aggregator to compute bag-level predic￾tions and aggregator loss Lagg (solid arrows). During backpropagation (dashed arrows), gradients from Lagg are detached to formulate t… view at source ↗
Figure 3
Figure 3. For bianry FGFR3 prediction task in BLCA – (a) validation AUC [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Ablations studies for FGFR3 prediction in BLCA patients. [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 30 canonical work pages

  1. [1]

    Artificial intelligence for digital and computational pathology

    Andrew H Song, Guillaume Jaume, Drew FK Williamson, Ming Y Lu, Anurag Vaidya, Tiffany R Miller, and Faisal Mahmood. Artificial intelligence for digital and computational pathology. Nature Reviews Bioengineering, 1(12):930–949, 2023

    2023

  2. [2]

    Attention-based deep multiple instance learning

    Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. InInternational Conference on Machine Learning, pages 2127–

    work page

  3. [3]

    Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J

    Gabriele Campanella, Matthew G. Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J. Busam, Edi Brogi, Victor E. Reuter, David S. Klim- stra, and Thomas J. Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images.Nature Medicine, 25(8):1301–1309, 2019

    work page 2019

  4. [4]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representations, 2021

    work page 2021

  5. [5]

    Towards a general-purpose foundation model for computational pathology

    Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850–862, 2024

    2024

  6. [6]

    A foundation model for clinical-grade computational pathology and rare cancers detection.Nature medicine, 30(10):2924–2935, 2024

    Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, et al. A foundation model for clinical-grade computational pathology and rare cancers detection.Nature medicine, 30(10):2924–2935, 2024

    work page 2024

  7. [7]

    Virchow2: Scaling self-supervised mixed magnification models in pathology

    Eric Zimmermann, Eugene Vorontsov, Julian Viret, Adam Casson, Michal Zele- chowski, George Shaikovski, Neil Tenenholtz, James Hall, David Klimstra, Razik Yousfi, et al. Virchow2: Scaling self-supervised mixed magnification models in pathology. arXiv preprint arXiv:2408.00738, 2024

    arXiv 2024

  8. [8]

    A whole-slide foundationmodelfordigitalpathologyfromreal-worlddata

    Hanwen Xu, Naoto Usuyama, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Naumann, Cliff Wong, Zelalem Gero, Javier González, Yu Gu, et al. A whole-slide foundationmodelfordigitalpathologyfromreal-worlddata. Nature, 630(8015):181– 188, 2024

    work page 2024

Show all 39 references
  1. [9]

    H-optimus-0: A foundational model for histopathology

    Bioptimus. H-optimus-0: A foundational model for histopathology. https:// github.com/bioptimus/releases/tree/main/models/h-optimus/v0, 2024. Ac- cessed: April 2025

    2024

  2. [10]

    M. Y. Lu, D. F. K. Williamson, T. Y. Chen, R. J. Chen, M. Barbieri, and F. Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering, 5(6):555–570, 2021

    2021

  3. [11]

    A benchmarking crisis in biomedical machine learning.Nature Medicine, pages 1–1, 2025

    Faisal Mahmood. A benchmarking crisis in biomedical machine learning.Nature Medicine, pages 1–1, 2025. 15

    2025

  4. [12]

    Validation of histopathology foundation models through whole slide image retrieval.Scientific Reports, 15(1):3990, 2025

    Saghir Alfasly, Ghazal Alabtah, Sobhan Hemati, Krishna Rani Kalari, Joaquin J Garcia, and HR Tizhoosh. Validation of histopathology foundation models through whole slide image retrieval.Scientific Reports, 15(1):3990, 2025

    2025

  5. [13]

    A clinical benchmark of public self-supervised pathology foundation models

    Gabriele Campanella, Shengjia Chen, Manbir Singh, Ruchika Verma, Silke Muehlst- edt, Jennifer Zeng, Aryeh Stock, Matt Croken, Brandon Veremis, Abdulkadir Elmas, et al. A clinical benchmark of public self-supervised pathology foundation models. Nature Communications, 16(1):3640, 2025

    2025

  6. [14]

    Transformer-based unsupervised contrastive learning for histopathological image classification.Medical image analysis, 81:102559, 2022

    Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Wei Yang, Jun- zhou Huang, and Xiao Han. Transformer-based unsupervised contrastive learning for histopathological image classification.Medical image analysis, 81:102559, 2022

    2022

  7. [15]

    Scaling vision transformers to gigapixel images via hierarchical self-supervised learning

    Richard J Chen, Chengkuan Chen, Yicong Li, Tiffany Y Chen, Andrew D Trister, Rahul G Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. InProceedings of the IEEE/CVF conference on computer vision and pattern...

    2022

  8. [16]

    Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging

    Shekoofeh Azizi, Laura Culp, Jan Freyberg, Basil Mustafa, Sebastien Baur, Simon Kornblith, Ting Chen, Nenad Tomasev, Jovana Mitrović, Patricia Strachan, et al. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging. Nature Biomedica...

    2023

  9. [17]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650–9660, 2021

    2021

  10. [18]

    Scaling self-supervised learning for histopathology with masked image modeling

    Alexandre Filiot, Ridouane Ghermi, Antoine Olivier, Paul Jacob, Lucas Fidon, Axel Camara, Alice Mac Kain, Charlie Saillard, and Jean-Baptiste Schiratti. Scaling self-supervised learning for histopathology with masked image modeling. medRxiv, pages 2023–07, 2023

    2023

  11. [19]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabba...

    2024

  12. [20]

    A visual-language foundation model for computational pathology.Nature Medicine, 30(3):863–874, 2024

    Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology.Nature Medicine, 30(3):863–874, 2024

    2024

  13. [21]

    Prism: A multi-modal generative foundation model for slide-level histopathology.arXiv preprint arXiv:2405.10254, 2024

    George Shaikovski, Adam Casson, Kristen Severson, Eric Zimmermann, Yi Kan Wang, Jeremy D Kunz, Juan A Retamero, Gerard Oakley, David Klimstra, Christopher Kanan, et al. Prism: A multi-modal generative foundation model for slide-level histopathology.arXiv preprint arXiv:2405.10...

    2024 arXiv

  14. [22]

    A vision–language foundation model for precision oncology.Nature, pages 1–10, 2025

    Jinxi Xiang, Xiyue Wang, Xiaoming Zhang, Yinghua Xi, Feyisope Eweje, Yijiang Chen, Yuchen Li, Colin Bergstrom, Matthew Gopaulchan, Ted Kim, et al. A vision–language foundation model for precision oncology.Nature, pages 1–10, 2025. 16

    2025

  15. [23]

    Multimodal histopathologic models stratify hormone receptor-positive early breast cancer

    Kevin M Boehm, Omar SM El Nahhas, Antonio Marra, Michele Waters, Justin Jee, Lior Braunstein, Nikolaus Schultz, Pier Selenica, Hannah Y Wen, Britta Weigelt, et al. Multimodal histopathologic models stratify hormone receptor-positive early breast cancer. Nature Communications, ...

    2025

  16. [24]

    Dual-stream multiple instance learning networkforwholeslide imageclassificationwithself-supervised contrastive learning

    Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning networkforwholeslide imageclassificationwithself-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318–14328, 2021

    2021

  17. [25]

    Deepsmile: contrastive self-supervised pre-training benefits msi and hrd classification directly from h&e whole-slide images in colorectal and breast cancer

    Yoni Schirris, Efstratios Gavves, Iris Nederlof, Hugo Mark Horlings, and Jonas Teuwen. Deepsmile: contrastive self-supervised pre-training benefits msi and hrd classification directly from h&e whole-slide images in colorectal and breast cancer. Medical Image Analysis, 79:102464, 2022

    2022

  18. [26]

    Transmil: Transformer based correlated multiple instance learning for whole slide image classification

    Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems, 34:2136–2147, 2021

    2021

  19. [27]

    Gindra, Emily J

    Yi Zheng, Rushin H. Gindra, Emily J. Green, Eric J. Burks, Margrit Betke, Jennifer E. Beane, and Vijaya B. Kolachalama. A graph-transformer for whole slide image classification.IEEE Transactions on Medical Imaging, 41(11):3003– 3015, 2022

    2022

  20. [28]

    Polydorides, Adam J

    Shengjia Chen, Gabriele Campanella, Abdulkadir Elmas, Aryeh Stock, Jennifer Zeng, Alexandros D. Polydorides, Adam J. Schoenfeld, Kuan lin Huang, Jane Houldsworth, Chad Vanderbilt, and Thomas J. Fuchs. Benchmarking embedding aggregation methods in computational pathology: A cli...

    2024

  21. [29]

    Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image clas- sification

    Honglin Li, Chenglu Zhu, Yunlong Zhang, Yuxuan Sun, Zhongyi Shui, Wenwei Kuang, Sunyi Zheng, and Lin Yang. Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image clas- sification. In Proceedings of the IEEE/CVF confer...

    2023

  22. [30]

    Beyond multiple instance learning: Full resolution all-in-memory end-to-end pathology slide modeling.arXiv preprint arXiv:2403.04865, 2024

    Gabriele Campanella, Eugene Fluder, Jennifer Zeng, Chad Vanderbilt, and Thomas J Fuchs. Beyond multiple instance learning: Full resolution all-in-memory end-to-end pathology slide modeling.arXiv preprint arXiv:2403.04865, 2024

    2024 arXiv

  23. [31]

    DEMO: A scalable artificial intelligence framework for rapid EGFR mutation screening in lung cancer

    Neeraj Kumar, Swaraj Nanda, Siddharth Singi, Gabriele Campanella, Thomas Fuchs, and Chad Vanderbilt. DEMO: A scalable artificial intelligence framework for rapid EGFR mutation screening in lung cancer. InGenAI for Health: Potential, Trust and Policy Compliance, 2024

    2024

  24. [32]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019

    2019

  25. [33]

    Sgdr: Stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2017

    2017

  26. [34]

    Oncogenicalterationsinadvancednsclc: amolecularsuper-highway

    Alex Friedlaender, Maurice Perol, Giuseppe Luigi Banna, Kaushal Parikh, and AlfredoAddeo. Oncogenicalterationsinadvancednsclc: amolecularsuper-highway. Biomarker Research, 12(1):24, 2024. 17

    2024

  27. [35]

    Oncogenic driver mutations in non-small cell lung cancer: Past, present and future

    Mathieu Chevallier, Maxime Borgeaud, Alfredo Addeo, and Alex Friedlaender. Oncogenic driver mutations in non-small cell lung cancer: Past, present and future. World Journal of Clinical Oncology, 12(4):217, 2021

    2021

  28. [36]

    Neal I Lindeman, Philip T Cagle, Mary Beth Beasley, Dhananjay Arun Chitale, Sanja Dacic, Giuseppe Giaccone, Robert Brian Jenkins, David J Kwiatkowski, Juan-Sebastian Saldivar, Jeremy Squire, et al. Molecular testing guideline for selectionoflungcancerpatientsforegfrandalktyros...

    2013

  29. [37]

    Neal I Lindeman, Philip T Cagle, Dara L Aisner, Maria E Arcila, Mary Beth Beasley, Eric H Bernicker, Carol Colasacco, Sanja Dacic, Fred R Hirsch, Keith Kerr, et al. Updated molecular testing guideline for the selection of lung cancer patients for treatment with targeted tyrosi...

    2018

  30. [38]

    Molecular pathology of non–small cell lung cancer: a practical guide.American journal of clinical pathology, 138(3):332–346, 2012

    Dara L Aisner and Carrie B Marshall. Molecular pathology of non–small cell lung cancer: a practical guide.American journal of clinical pathology, 138(3):332–346, 2012

    2012

  31. [39]

    Role of fgfr3 in bladder cancer: Treatment landscape and future challenges.Cancer Treatment Reviews, 115:102530, 2023

    Claudia Maria Ascione, Fabiana Napolitano, Daniela Esposito, Alberto Servetto, Stefania Belli, Antonio Santaniello, Sarah Scagliarini, Felice Crocetto, Roberto Bianco, and Luigi Formisano. Role of fgfr3 in bladder cancer: Treatment landscape and future challenges.Cancer Treatm...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.