Pith. sign in

REVIEW 3 major objections 4 minor 48 references

Scalable and Loosely-Coupled Multimodal Deep Learning for Breast Cancer Subtyping

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper argues that combining CNV, clinical records, and a dual image-plus-graph representation of whole-slide images, fused by weighted logits with bias, outperforms existing methods at PAM50 breast cancer subtyping.

desk verdict Useful engineering result with a plausible SOTA claim, but the paper never states whether the 10-fold CV is patient-level, and without that detail the headline numbers are not fully interpretable. read the letter →

arxiv 2509.03408 v1 pith:BEJIIHFU submitted 2025-09-03 cs.CV cs.LG

Mohammed Amer , Mohamed A. Suliman , Tu Bui , Nuria Garcia , Serban Georgescu This is my paper
classification cs.CVcs.LG
keywords multimodaldeeplearningbreastcancersubtypingPAM50whole-slideimagesgraphneuralnetworkslatefusioncopynumbervariationclinicalrecords
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 tries to show that breast cancer molecular subtyping improves when several clinically available data types are combined in a flexible, loosely coupled way: copy-number variation, structured clinical records, and whole-slide histology images represented twice—once as patch images and once as tissue graphs. It claims the combination, merged by a weighted-logits late-fusion rule with a bias term, outperforms existing fusion strategies on the PAM50 four-subtype task, reaching 78.13% accuracy and 0.9153 macro-AUC. The reason this matters is practical: hospitals differ in which tests they can run, so a system that can add or drop modalities without retraining the rest would be easier to deploy in real clinical workflows. The paper also argues the dual slide representation itself is a source of improvement, not just the fusion rule.

What carries the argument

The load-bearing object is the dual-based whole-slide image representation. One branch treats each slide as a collection of tissue-rich patches scored by a pretrained CNN; the other converts the same slide into a graph whose nodes are spatially and compositionally clustered patch features and whose edges come from Delaunay triangulation, then processes that graph with graph-convolution layers. The second load-bearing object is the fusion rule: the fused logit for each class is a weighted sum of the per-modality logits for that class plus a trainable bias, followed by a task-specific nonlinearity. Because each modality model is trained independently first, the fusion step is lightweight and m

What would settle it

Run the exact pipeline with patient-level 10-fold cross-validation, where each patient's slides and patches appear in only one fold, and compare accuracy and macro-AUC to the reported 78.13% and 0.9153. Also report the variation across folds; if the lead over the best baseline shrinks to near zero or the absolute scores drop substantially, the central outperformance claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a multimodal model built from four components—CNV, clinical/EHR data, a CNN over WSI patches, and a graph neural network over a cell-graph version of the same slides—classifies PAM50 breast cancer subtypes better than any single modality and better than established fusion methods. The paper's full model reaches 78.13% accuracy and 0.9153 macro-AUC on the 977-patient TCGA-BRCA set, and the weighted-logits-with-bias fusion is reported as the best or second-best method in almost every modality combination tested. A supporting claim is that the graph representation carries complementary information: adding it to the image representation improves accuracy by roughly one

Load-bearing premise

The whole comparison is valid only if the 10-fold cross-validation keeps all slides and patches from the same patient in the same fold; the paper never states that, and without it the reported accuracy and AUC could be inflated by seeing the same patient in both training and test.

Editorial extensions

If this is right

  • Combining all four components raises accuracy from the best single modality's 70.43% to 78.13% and macro-AUC from 0.8522 to 0.9153.
  • The graph view of a slide adds about 1% accuracy and 0.02 macro-AUC on top of the image view in multimodal configurations, indicating tissue topology carries signal beyond patch appearance.
  • The weighted-logits fusion with a bias term beats more complex alternatives such as transformer-based fusion, meta-learning, and intermediate fusion in most comparisons, so a cheap late-fusion rule may be sufficient for this clinical task.
  • Because modality models are decoupled, a clinical site can use only the modalities it has and still obtain predictions, and new modalities can be added later with only fusion retraining.
  • The same pipeline is designed to transfer to other cancer subtyping tasks that have CNV, EHR, and histology data, though the paper demonstrates this only for breast cancer.

Reading between the lines

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

  • The reported 78.13% and 0.9153 numbers rest on the unstated assumption that the 10-fold cross-validation is split by patient; a patient-level rerun is a direct way to test how much of the gain is real.
  • Because the fusion weights are non-normalized and trained after the single-modality models, the method is likely sensitive to logit scale differences across modalities; temperature-scaling each modality before fusion could change the reported gains.
  • The dual WSI idea suggests a natural testable extension: applying the same graph pipeline to other slide-level tasks such as HER2 status or survival to isolate how much of the gain comes from tissue topology versus the fusion rule.
  • Interpretability weights imply the graph branch dominates the fusion at 28% attribution; if that is stable, graph-construction choices such as node clustering thresholds and Delaunay distance may influence downstream fusion accuracy as much as the classification heads.
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

3 major / 4 minor

Summary. The paper proposes a scalable, loosely coupled multimodal framework for PAM50 breast cancer subtyping, integrating CNV, clinical/EHR data, and a dual WSI representation (image-based and graph-based). The method uses independent single-modality models whose logits are combined by a trainable weighted late-fusion strategy, optionally with a bias term. Experiments on 977 TCGA-BRCA patients are evaluated with 10-fold cross-validation and compared against six established fusion baselines across several modality combinations. The authors report a best four-modality accuracy of 78.13% and macro-AUC of 0.9153, claiming state-of-the-art performance.

Significance. If the reported evaluation is valid, the paper makes a useful contribution: the dual WSI representation (CNN features plus graph-based cell/tissue features) appears to add complementary information, and the loosely coupled design allows modalities to be added or removed without retraining existing branches. The authors also benchmark against a broader set of fusion baselines than is common and include interpretability analyses. However, the central 'outperforms SOTA' claim is currently not fully supported because key procedural details of the cross-validation — most importantly the splitting unit — are not stated, and no statistical significance or variance estimates are provided. The strengths of the system design and benchmarking are real, but the evaluation needs to be tightened before the headline claim can be accepted.

major comments (3)
  1. [Section 4 (Experiments)] The evaluation is described only as '10-fold cross-validation,' with no statement of the splitting unit. This is critical because the WSI model is trained on patches that inherit patient labels (Section 3.2.1), each patient contributes multiple WSIs and up to 50 tissue-rich patches, and the graph model constructs per-slide graphs from patient tissue. If the folds are not split by patient, patches or graphs from the same patient can appear in both training and test folds, allowing the model to memorize patient-specific tissue patterns and inflating the reported accuracy and macro-AUC. Please state explicitly that the split is at the patient level, report the number of patients per fold, and describe how patch extraction and graph construction are confined to the training fold (i.e., test-slide information is not used during preprocessing or model selection).
  2. [Tables 2 and 3; Section 5 (Results and Discussion)] No standard deviations, confidence intervals, or statistical significance tests are reported. In the four-modality setting, the gain over Simple Ensemble is 0.17% accuracy and the macro-AUC is tied at 0.9153. Without fold-wise results and paired comparisons (e.g., McNemar's test for accuracy, DeLong's test for correlated AUCs), the abstract's claim that the method 'outperforms state-of-the-art methods' is not established. Please report per-fold metrics, standard deviations, and appropriate paired tests for the main comparisons.
  3. [Section 3.1, Eq. (1); Section 4 (Experiments)] The fusion weights w_ij and bias b_j in Eq. (1) are trainable, and the text says a 'second training phase' optimizes fusion. However, it is not specified whether this fusion training is nested inside each cross-validation fold or performed once on the full dataset. If the fusion weights are learned on all patients before evaluating held-out folds, the fold boundaries are not respected and the reported results are no longer properly out-of-sample. Please clarify the exact protocol and, if necessary, modify it so that the fusion weights are trained only on training folds.
minor comments (4)
  1. [Section 3.2.1] The patch selection procedure ('top 50 tissue-rich patches per patient') is not fully specified; please state how tissue-richness is scored and whether this selection is performed per WSI or per patient and within each training fold.
  2. [Tables 2 and 3] In the four-modality row, WLB ties with SE at 0.9153 macro-AUC. The text in Section 5 says the method 'surpasses SOTA in nearly all cases'; this should be phrased more precisely to acknowledge the tie.
  3. [Figure 1] The in-text references to 'Fig. 1i (A)' and 'Fig. 1i (B)' are confusing because the figure is placed later and contains multiple panels. Please label the panels clearly and refer to them consistently.
  4. [General] No code or data availability statement is provided. Given the sensitivity of the evaluation protocol to the split unit, releasing the patient-level fold assignments or code would materially help reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper is an empirical benchmark with out-of-sample evaluation and no load-bearing self-citation.

full rationale

The paper makes no derivation from first principles; its central claim is an empirical performance comparison. Equation (1) defines the proposed weighted-logits fusion, and the fusion weights w_ij and biases b_j are trained on the training folds of the 10-fold cross-validation and evaluated on held-out folds, so the reported accuracy/AUC are out-of-sample results rather than re-statements of fitted constants. No single-modality model or fused result is defined in terms of the target metric or the final SOTA comparison. The references to prior work (e.g., Lu et al. 2021 for patch extraction, SlideGraph+ for graph construction, pre-trained backbones) are external and do not smuggle in the paper's conclusion. There are no self-citations invoked as load-bearing support, and no uniqueness theorem or ansatz is imported from the authors' own prior work. The main validity concern is procedural, not circular: Section 4 states only "Evaluation uses 10-fold cross-validation" without specifying patient-level splitting, which could permit same-patient leakage for the WSI modality; however, that is a potential data-leakage/correctness issue and does not amount to the paper's derivation reducing to its own inputs. Accordingly, no circular step is present.

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

The paper introduces no new physical or conceptual entities. The free parameters are standard trainable weights and hand-chosen hyperparameters of the machine learning pipeline, all openly described. The most important latent assumption is the unstated patient-level split, which the central performance claims rely on.

free parameters (4)
  • Fusion weights w_ij and bias b_j = Learned from training folds
    These are the trainable parameters of the weighted logits fusion, Eq. (1). The central claim that the fusion outperforms baselines depends on these values being well estimated from data.
  • Graph construction thresholds = spatial threshold=2000, gamma=0.001, linkage cut-off=0.8, Delaunay threshold=4000
    Hand-chosen constants in Section 3.2.2 that determine node clustering and edge connectivity in the WSI graph. Different values would change the graph and likely the model's performance.
  • WSI patch selection hyperparameters = tissue area >5%, saturation threshold=20, median blur kernel=7, top-50 patches
    Hand-chosen values in Section 3.2.1 controlling which patches are used to train the WSI image model and graph model.
  • Model architecture hyperparameters = GNN L=3, dims [32,16,8], lr=1e-4, dropout=0.2; SNN 8192-2048; MLP 128-64
    Selected for each modality; results in Tables 1-3 depend on these choices, and the paper does not report sensitivity to them.
assumptions (4)
  • domain assumption TCGA-BRCA PAM50 labels from Netanely et al. (2016) are accurate ground truth.
    Section 4 states labels are taken from Netanely et al.; the evaluation treats these as error-free, but subtype calls from expression/methylation may be imperfect.
  • ad hoc to paper The 10-fold cross-validation split is patient-level (no patient overlap across folds).
    Section 4 describes 10-fold CV but does not explicitly state patient-level splitting. If not patient-level, patch/WSI leakage inflates performance.
  • domain assumption Public TCGA-BRCA data and cBioPortal CNV/clinical data are correctly preprocessed and aligned across modalities.
    Section 4 uses TCGA WSIs, cBioPortal CNV and clinical data, and Netanely PAM50 labels; matching patients across sources is assumed correct.
  • domain assumption Pre-trained models (InceptionV3, Dinov2, HoverNet, BCSS model) are appropriate feature extractors without domain-shift issues.
    Section 3.2 uses pre-trained models on natural or other histology images; the paper does not assess whether representation quality is sufficient, but the whole pipeline depends on these features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scalable and Loosely-Coupled Multimodal Deep Learning for Breast Cancer Subtyping." pith.science (2026). https://pith.science/paper/BEJIIHFU

@misc{pith2026250903408,
  author       = {Pith},
  title        = {Pith review of: Scalable and Loosely-Coupled Multimodal Deep Learning for Breast Cancer Subtyping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BEJIIHFU}},
  note         = {Machine review of arXiv:2509.03408}
}
read the original abstract

Healthcare applications are inherently multimodal, benefiting greatly from the integration of diverse data sources. However, the modalities available in clinical settings can vary across different locations and patients. A key area that stands to gain from multimodal integration is breast cancer molecular subtyping, an important clinical task that can facilitate personalized treatment and improve patient prognosis. In this work, we propose a scalable and loosely-coupled multimodal framework that seamlessly integrates data from various modalities, including copy number variation (CNV), clinical records, and histopathology images, to enhance breast cancer subtyping. While our primary focus is on breast cancer, our framework is designed to easily accommodate additional modalities, offering the flexibility to scale up or down with minimal overhead without requiring re-training of existing modalities, making it applicable to other types of cancers as well. We introduce a dual-based representation for whole slide images (WSIs), combining traditional image-based and graph-based WSI representations. This novel dual approach results in significant performance improvements. Moreover, we present a new multimodal fusion strategy, demonstrating its ability to enhance performance across a range of multimodal conditions. Our comprehensive results show that integrating our dual-based WSI representation with CNV and clinical health records, along with our pipeline and fusion strategy, outperforms state-of-the-art methods in breast cancer subtyping.

Figures

Figures reproduced from arXiv: 2509.03408 by the authors.

Figure 1
Figure 1. (i) The model structure. (ii) Graph generation steps. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 42 canonical work pages

  1. [1]

    Representation learning of histopathology images using graph neural networks

    Mohammed Adnan, Shivam Kalra, and Hamid R Tizhoosh. Representation learning of histopathology images using graph neural networks. pages 988--989, 2020

    work page 2020

  2. [2]

    Estrogen and progesterone receptor testing in breast cancer: Asco/cap guideline update

    Kimberly H Allison, M Elizabeth H Hammond, Mitchell Dowsett, Shannon E McKernin, Lisa A Carey, Patrick L Fitzgibbons, Daniel F Hayes, Sunil R Lakhani, Mariana Chavez-MacGregor, Jane Perlmutter, et al. Estrogen and progesterone receptor testing in breast cancer: Asco/cap guideline update. Journal of Clinical Oncology, 38 0 (12): 0 1346--1366, 2020

    work page 2020

  3. [3]

    Structured crowdsourcing enables convolutional segmentation of histology images

    Mohamed Amgad, Habiba Elfandy, Hagar Hussein, Lamees A Atteya, Mai AT Elsebaie, Lamia S Abo Elnasr, Rokia A Sakr, Hazem SE Salem, Ahmed F Ismail, Anas M Saad, et al. Structured crowdsourcing enables convolutional segmentation of histology images. Bioinformatics, 35 0 (18): 0 3461--3467, 2019

    work page 2019

  4. [4]

    Big-graph: Brain imaging genetics by graph neural network

    Javad Ansarifar, Iliana I Karipidis, Manish Saggar, and David S Hong. Big-graph: Brain imaging genetics by graph neural network. 2022

    work page 2022

  5. [5]

    Cell-graph mining for breast tissue modeling and classification

    Cagatay Bilgin, Cigdem Demir, Chandandeep Nagi, and Bulent Yener. Cell-graph mining for breast tissue modeling and classification. pages 5311--5314, 2007

    work page 2007

  6. [6]

    Obesity and breast cancer: progress to understanding the relationship

    Kristy A Brown and Evan R Simpson. Obesity and breast cancer: progress to understanding the relationship. Cancer research, 70 0 (1): 0 4--7, 2010

    work page 2010

  7. [7]

    Kristy A Brown, Kerry J McInnes, Nicole I Hunger, Jonathan S Oakhill, Gregory R Steinberg, and Evan R Simpson. Subcellular localization of cyclic amp-responsive element binding protein-regulated transcription coactivator 2 provides a link between obesity and breast cancer in postmenopausal women. Cancer research, 69 0 (13): 0 5392--5399, 2009

    work page 2009

  8. [8]

    The cbio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data

    Ethan Cerami, Jianjiong Gao, Ugur Dogrusoz, Benjamin E Gross, Selcuk Onur Sumer, B \"u lent Arman Aksoy, Anders Jacobsen, Caitlin J Byrne, Michael L Heuer, Erik Larsson, et al. The cbio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. Cancer discovery, 2 0 (5): 0 401--404, 2012

    work page 2012

Show all 48 references
  1. [9]

    There is more than one kind of myofibroblast: analysis of cd34 expression in benign, in situ, and invasive breast lesions

    H Chauhan, A Abraham, JRA Phillips, JH Pringle, RA Walker, and JL Jones. There is more than one kind of myofibroblast: analysis of cd34 expression in benign, in situ, and invasive breast lesions. Journal of clinical pathology, 56 0 (4): 0 271--276, 2003

    2003

  2. [10]

    Pan-cancer integrative histology-genomic analysis via multimodal deep learning

    Richard J Chen, Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Jana Lipkova, Zahra Noor, Muhammad Shaban, Maha Shady, Mane Williams, Bumjin Joo, et al. Pan-cancer integrative histology-genomic analysis via multimodal deep learning. Cancer Cell, 40 0 (8): 0 865--878, 2022

    2022

  3. [11]

    Expression of cluster of differentiation 34 and vascular endothelial growth factor in breast cancer, and their prognostic significance

    Zhanhong Chen, Shenhua Xu, Weizhen Xu, JIAN Huang, GU Zhang, LEI Lei, Xiying Shao, and Xiaojia Wang. Expression of cluster of differentiation 34 and vascular endothelial growth factor in breast cancer, and their prognostic significance. Oncology Letters, 10 0 (2): 0 723--729, 2015

    2015

  4. [12]

    Diagnostic significance of the immunoexpression of cd34 and smooth muscle cell actin in benign and malignant tumors of the breast

    Anca Maria Cimpean, M Raica, and D Narita. Diagnostic significance of the immunoexpression of cd34 and smooth muscle cell actin in benign and malignant tumors of the breast. Rom J Morphol Embryol, 46 0 (2): 0 123--129, 2005

    2005

  5. [13]

    Principal neighbourhood aggregation for graph nets

    Gabriele Corso, Luca Cavalleri, Dominique Beaini, Pietro Li \`o , and Petar Veli c kovi \'c . Principal neighbourhood aggregation for graph nets. Advances in Neural Information Processing Systems, 33: 0 13260--13271, 2020

    2020

  6. [14]

    Tumour heterogeneity and resistance to cancer therapies

    Ibiayi Dagogo-Jack and Alice T Shaw. Tumour heterogeneity and resistance to cancer therapies. Nature reviews Clinical oncology, 15 0 (2): 0 81--94, 2018

    2018

  7. [15]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  8. [16]

    Pannuke dataset extension, insights and baselines

    Jevgenij Gamper, Navid Alemi Koohbanani, Ksenija Benes, Simon Graham, Mostafa Jahanifar, Syed Ali Khurram, Ayesha Azam, Katherine Hewitt, and Nasir Rajpoot. Pannuke dataset extension, insights and baselines. arXiv preprint arXiv:2003.10778, 2020

    2003 arXiv

  9. [17]

    Estrogen receptor status, determined by immunohistochemistry, as a predictor of the recurrence of stage i endometrial carcinoma

    Paola A Gehrig, Linda Van Le, Babatunde Olatidoye, and Joseph Geradts. Estrogen receptor status, determined by immunohistochemistry, as a predictor of the recurrence of stage i endometrial carcinoma. Cancer, 86 0 (10): 0 2083--2089, 1999

  10. [18]

    Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images

    Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical image analysis, 58: 0 101563, 2019

    2019

  11. [19]

    Multimodal fusion with deep neural networks for leveraging ct imaging and electronic health record: a case-study in pulmonary embolism detection

    Shih-Cheng Huang, Anuj Pareek, Roham Zamanian, Imon Banerjee, and Matthew P Lungren. Multimodal fusion with deep neural networks for leveraging ct imaging and electronic health record: a case-study in pulmonary embolism detection. Scientific reports, 10 0 (1): 0 22147, 2020

    2020

  12. [20]

    Self-normalizing neural networks

    G \"u nter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. Advances in neural information processing systems, 30, 2017

    2017

  13. [21]

    Topological feature extraction and visualization of whole slide images using graph neural networks

    Joshua Levy, Christian Haudenschild, Clark Barwick, Brock Christensen, and Louis Vaickus. Topological feature extraction and visualization of whole slide images using graph neural networks. pages 285--296, 2020

    2020

  14. [22]

    Mmgk: Multimodality multiview graph representations and knowledge embedding for mild cognitive impairment diagnosis

    Jin Liu, Hao Du, Rui Guo, Harrison X Bai, Hulin Kuang, and Jianxin Wang. Mmgk: Multimodality multiview graph representations and knowledge embedding for mild cognitive impairment diagnosis. IEEE Transactions on Computational Social Systems, 2022 a

    2022

  15. [23]

    A hybrid deep learning model for predicting molecular subtypes of human breast cancer using multimodal data

    T Liu, J Huang, T Liao, R Pu, S Liu, and Y Peng. A hybrid deep learning model for predicting molecular subtypes of human breast cancer using multimodal data. Irbm, 43 0 (1): 0 62--74, 2022 b

    2022

  16. [24]

    Roberta: A robustly optimized bert pretraining approach

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019

    1907 arXiv

  17. [25]

    Feature driven local cell graph (fedeg): predicting overall survival in early stage lung cancer

    Cheng Lu, Xiangxue Wang, Prateek Prasanna, German Corredor, Geoffrey Sedor, Kaustav Bera, Vamsidhar Velcheti, and Anant Madabhushi. Feature driven local cell graph (fedeg): predicting overall survival in early stage lung cancer. pages 407--416, 2018

    2018

  18. [26]

    Data-efficient and weakly supervised computational pathology on whole-slide images

    Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5 0 (6): 0 555--570, 2021

    2021

  19. [27]

    Capturing cellular topology in multi-gigapixel pathology images

    Wenqi Lu, Simon Graham, Mohsin Bilal, Nasir Rajpoot, and Fayyaz Minhas. Capturing cellular topology in multi-gigapixel pathology images. pages 260--261, 2020

    2020

  20. [28]

    Slidegraph+: Whole slide image level graphs to predict her2 status in breast cancer

    Wenqi Lu, Michael Toss, Muhammad Dawood, Emad Rakha, Nasir Rajpoot, and Fayyaz Minhas. Slidegraph+: Whole slide image level graphs to predict her2 status in breast cancer. Medical Image Analysis, 80: 0 102486, 2022

    2022

  21. [29]

    Expression and methylation patterns partition luminal-a breast tumors into distinct prognostic subgroups

    Dvir Netanely, Ayelet Avraham, Adit Ben-Baruch, Ella Evron, and Ron Shamir. Expression and methylation patterns partition luminal-a breast tumors into distinct prognostic subgroups. Breast Cancer Research, 18: 0 1--16, 2016

    2016

  22. [30]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth \'e e Darcet, Th \'e o Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  23. [31]

    Supervised risk predictor of breast cancer based on intrinsic subtypes

    Joel S Parker, Michael Mullins, Maggie CU Cheang, Samuel Leung, David Voduc, Tammi Vickery, Sherri Davies, Christiane Fauron, Xiaping He, Zhiyuan Hu, et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. Journal of clinical oncology, 27 0 (8): 0 1160, 2009

    2009

  24. [32]

    Hact-net: A hierarchical cell-to-tissue graph neural network for histopathological image classification

    Pushpak Pati, Guillaume Jaume, Lauren Alisha Fernandes, Antonio Foncubierta-Rodr \' guez, Florinda Feroce, Anna Maria Anniciello, Giosue Scognamiglio, Nadia Brancati, Daniel Riccio, Maurizio Di Bonito, et al. Hact-net: A hierarchical cell-to-tissue graph neural network for his...

    2020

  25. [33]

    High timm17a expression is associated with adverse pathological and clinical outcomes in human breast cancer

    Mohamed Salhab, Neill Patani, Wen Jiang, and Kefah Mokbel. High timm17a expression is associated with adverse pathological and clinical outcomes in human breast cancer. Breast cancer, 19: 0 153--160, 2012

    2012

  26. [34]

    Creb-regulated transcription co-activator family stimulates promoter ii-driven aromatase expression in preadipocytes

    Nirukshi U Samarajeewa, Maria M Docanto, Evan R Simpson, and Kristy A Brown. Creb-regulated transcription co-activator family stimulates promoter ii-driven aromatase expression in preadipocytes. Hormones and Cancer, 4: 0 233--241, 2013

    2013

  27. [35]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014

    2014 arXiv

  28. [36]

    Multimodal deep learning for biomedical data fusion: a review

    S \"o ren Richard Stahlschmidt, Benjamin Ulfenborg, and Jane Synnergren. Multimodal deep learning for biomedical data fusion: a review. Briefings in bioinformatics, 23 0 (2): 0 bbab569, 2022

    2022

  29. [37]

    Multimodal data fusion for cancer biomarker discovery with deep learning

    Sandra Steyaert, Marija Pizurica, Divya Nagaraj, Priya Khandelwal, Tina Hernandez-Boussard, Andrew J Gentles, and Olivier Gevaert. Multimodal data fusion for cancer biomarker discovery with deep learning. Nature machine intelligence, 5 0 (4): 0 351--362, 2023

    2023

  30. [38]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. pages 3319--3328, 2017

    2017

  31. [39]

    Rethinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. pages 2818--2826, 2016

    2016

  32. [40]

    Fusionbench: A comprehensive benchmark of deep model fusion

    Anke Tang, Li Shen, Yong Luo, Han Hu, Bo Du, and Dacheng Tao. Fusionbench: A comprehensive benchmark of deep model fusion. arXiv preprint arXiv:2406.03280, 2024

    2024

  33. [41]

    Her2 immunohistochemical scores provide prognostic information for patients with her2-type invasive breast cancer

    Yi-Fang Tsai, Ling-Ming Tseng, Pei-Ju Lien, Chih-Yi Hsu, Yen-Shu Lin, Kuang-Liang King, Yu-Ling Wang, Ta-Chung Chao, Chun-Yu Liu, Jen-Hwey Chiu, et al. Her2 immunohistochemical scores provide prognostic information for patients with her2-type invasive breast cancer. Histopatho...

    2019

  34. [42]

    Lung cancer subtype diagnosis by fusing image-genomics data and hybrid deep networks

    Xingze Wang, Guoxian Yu, Zhongmin Yan, Lin Wan, Wei Wang, and Lizhen Cui. Lung cancer subtype diagnosis by fusing image-genomics data and hybrid deep networks. IEEE/ACM transactions on computational biology and bioinformatics, 20 0 (1): 0 512--523, 2021

    2021

  35. [43]

    Redefining breast cancer subtypes to guide treatment prioritization and maximize response: Predictive biomarkers across 10 cancer therapies

    Denise M Wolf, Christina Yau, Julia Wulfkuhle, Lamorna Brown-Swigart, Rosa I Gallagher, Pei Rong Evelyn Lee, Zelos Zhu, Mark J Magbanua, Rosalyn Sayaman, Nicholas O’Grady, et al. Redefining breast cancer subtypes to guide treatment prioritization and maximize response: Predict...

    2022

  36. [44]

    Heterogeneous model reuse via optimizing multiparty multiclass margin

    Xi-Zhu Wu, Song Liu, and Zhi-Hua Zhou. Heterogeneous model reuse via optimizing multiparty multiclass margin. In International Conference on Machine Learning, pages 6840--6849. PMLR, 2019

    2019

  37. [45]

    Quantitative proteomics study of breast cancer cell lines isolated from a single patient: discovery of timm17a as a marker for breast cancer

    Xiaoen Xu, Meng Qiao, Yang Zhang, Yinghua Jiang, Ping Wei, Jun Yao, Bo Gu, Yaqi Wang, Jing Lu, Zhigang Wang, et al. Quantitative proteomics study of breast cancer cell lines isolated from a single patient: discovery of timm17a as a marker for breast cancer. Proteomics, 10 0 (7...

    2010

  38. [46]

    The impact of timm17a on aggressiveness of human breast cancer cells

    Xiaomei Yang, Yang Si, Tao Tao, Tracey A Martin, Shan Cheng, Hefen Yu, Jinyao Li, Junqi He, and Wen G Jiang. The impact of timm17a on aggressiveness of human breast cancer cells. Anticancer research, 36 0 (3): 0 1237--1241, 2016

    2016

  39. [47]

    Triple-negative breast cancer molecular subtyping and treatment progress

    Li Yin, Jiang-Jie Duan, Xiu-Wu Bian, and Shi-cang Yu. Triple-negative breast cancer molecular subtyping and treatment progress. Breast Cancer Research, 22: 0 1--13, 2020

    2020

  40. [48]

    Analysis of tumor genomic pathway alterations using broad-panel next-generation sequencing in surgically resected lung adenocarcinoma

    Jian Zhou, Francisco Sanchez-Vega, Raul Caso, Kay See Tan, Whitney S Brandt, Gregory D Jones, Shi Yan, Prasad S Adusumilli, Matthew Bott, James Huang, et al. Analysis of tumor genomic pathway alterations using broad-panel next-generation sequencing in surgically resected lung ...

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.