Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

GLAM: Geometry-Guided Local Alignment for Multi-View VLP in Mammography

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read GLAM, a vision-language model for mammography, aligns each local patch with its entire corresponding anatomical slice in the other view, and this geometry-aware alignment lets it outperform prior models in zero-shot, low-data, and fine-tuni

desk verdict A solid geometry-guided local alignment method for mammography VLP with consistent gains; the core geometry assumption is approximate and the evaluation lacks uncertainty estimates, but the central claim is defensible and the paper deserves review. read the letter →

arxiv 2509.10344 v1 pith:WLJLBTN6 submitted 2025-09-12 cs.CV cs.AIcs.LG

Yuexi Du , Lihui Chen , Nicha C. Dvornek This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords mammographyvision-languagepre-trainingcontrastivelearningmulti-viewalignmentgeometryguidanceCLIPbreastcancerscreeningfoundationmodel
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 GLAM, a contrastive vision-language pretraining method for mammograms that uses the known imaging geometry of the two standard views: craniocaudal (CC) and mediolateral oblique (MLO). Because a region of interest occupies the same anterior-posterior slice in both views, the model aligns each local super-patch in one view to the entire corresponding slice in the other view, rather than to a single patch. This local alignment is added to global contrastive losses that pull the two views together and align each view with its paired radiology report. Pretrained on EMBED, GLAM consistently outperforms prior mammography and medical CLIP models across zero-shot, linear probing, fine-tuning, and low-data settings on EMBED, VinDr, and RSNA-Mammo, showing that encoding breast projection geometry improves the transferability of mammography foundation models.

What carries the argument

The core mechanism is the geometry-guided patch-to-slice alignment along the anterior-posterior (AP) axis. After removing the pectoral region and rotating the MLO image so that the chest-to-nipple axis is parallel to the AP axis, the two views share a common AP coordinate. A query super-patch in one view attends over the entire AP slice in the other view via a multi-head cross-attention module, yielding a weighted-sum positive embedding that respects breast deformation. A contrastive loss with position-based and patient-based negatives forces the model to learn true anatomical correspondence, and this local loss is added to a global contrastive loss aligning the two views' global embeddings

What would settle it

A concrete test: take a set of mammogram pairs with expert annotations of the same lesion in both views and measure the cross-attention weight that GLAM assigns to the true corresponding AP slice in the opposite view versus other slices. If the model's attention at the true slice is not significantly higher than chance or than a model trained without geometry guidance, then the AP-correspondence assumption is not actually being learned. Alternatively, artificially shear or rotate one view after preprocessing and observe whether zero-shot multi-view accuracy drops steeply; a steep, monotonic dr

Watch

Extended reading notes

Core claim

The paper's central claim is that multi-view correspondence in mammograms is best modeled as patch-to-slice rather than patch-to-patch: because the CC and MLO views are projections of the same 3D breast along different angles, a patch in one view corresponds not to a unique point but to a whole slice along the anterior-posterior axis in the other view. GLAM uses this geometric constraint in a self-supervised contrastive objective, where each super-patch queries the opposite view's AP slice via cross-attention, producing a cross-view positive embedding; all other positions and same-position patches from other patients act as negatives. Combined with global multi-view and image-text contrastiv

Load-bearing premise

The load-bearing premise is that, after pectoral removal and rotation, patches at the same anterior-posterior position in the CC and MLO views correspond to the same breast tissue; breast deformation, patient positioning, and imperfect rotation can break this correspondence, in which case the local alignment loss would contrastively pull together unrelated tissue regions.

Editorial extensions

If this is right

  • Out-of-domain transfer improves: GLAM is the best or second-best method on 10 of 12 metrics on VinDr and RSNA-Mammo, indicating that geometry-guided local alignment generalizes beyond the pretraining distribution.
  • Low-data efficiency increases: with only 1% of EMBED training data, GLAM's linear probe outperforms almost all baselines trained on 100% of the data.
  • Multi-view zero-shot prediction gains from the learned correspondence: GLAM improves by about 2.5% balanced accuracy on BI-RADS when both views are used, while baselines show little or no such improvement.
  • Ablation results show that the geometry constraint itself, not just extra contrastive computation, is responsible for the gains: removing it or replacing AP-slice alignment with full cross-attention lowers performance.
  • The pretrained encoder can serve as a backbone for future multi-view fusion methods and dense multimodal tasks, as the authors state in their plans.

Reading between the lines

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

  • The patch-to-slice alignment idea could transfer to other multi-view imaging modalities where a known anatomical axis defines correspondence, such as two X-ray views of the same joint; if the geometry is known, similar slice-based alignment may improve representation learning without dense registration.
  • The same-position cross-patient negatives likely act as a hard-negative regularizer that discourages the model from relying on patient-identity cues; this effect could be studied in isolation on other medical imaging tasks with paired views.
  • Because GLAM synthesizes reports from tabular data, the text stream is relatively simple; a dataset with free-text radiology reports might magnify the benefit of the image-text contrastive loss, allowing a cleaner test of how much of the gain comes from language supervision versus geometry-guided local alignment.
  • The method's sensitivity to the pectoral-removal and rotation preprocessing is untested; a systematic degradation study that measures downstream accuracy as a function of induced AP-axis misalignment would quantify how heavily the result rests on the geometric preprocessing being accurate.
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 presents GLAM, a contrastive visual-language pre-training framework for multi-view mammography. It combines global multi-view and image–text contrastive losses with a novel geometry-guided local alignment loss: after pectoral removal and rotation, patches in the CC and MLO views are grouped into AP-axis slices, and each patch is contrasted with a cross-attention aggregation of the corresponding slice in the other view. The model is pre-trained on EMBED and evaluated on EMBED, VinDr, and RSNA-Mammo for BI-RADS, density, and cancer classification under zero-shot, linear-probing, and full fine-tuning settings. The authors report consistent improvements over several CLIP-style baselines and provide qualitative attention visualizations.

Significance. If the reported gains are robust, GLAM is a meaningful contribution to mammography VLP: it is one of the largest mammography CLIP models, it makes explicit use of the multi-view imaging geometry, and it ships code, which supports reproducibility. The evaluation covers multiple datasets and settings, and the method is clearly described. However, the central geometric assumption—that AP-axis slices correspond across CC and MLO views after preprocessing—is not directly validated, and the empirical advantages are reported without error bars or significance tests. The result is therefore promising but not yet fully established.

major comments (3)
  1. [Sec. 2.2, Eq. (3), and pre-processing paragraph] The core mechanism is the AP-slice correspondence after pectoral removal and rotation. The paper itself states that residual AP-axis misalignment can remain in extreme cases and is only softened by random affine augmentation. Because the breast is compressed and deformed differently in CC and MLO, residual offsets larger than the super-patch size (≈57 px for M=81 at 518 px input) will make L_local in Eq. (4) pull together tissue regions that are not anatomically matched. The ablation in Tab. 5 shows that the local loss helps, but it does not demonstrate that the specific geometric correspondence is correct; an auxiliary contrastive task could yield similar gains. Please add a quantitative evaluation of AP-alignment accuracy after preprocessing (e.g., using annotated landmarks or registration) and compare against a control that breaks the geometry, such as a fixed random slice corresponde
  2. [Tables 1–3, Experimental results] The paper reports only point estimates; no standard deviations, confidence intervals, or significance tests are provided. Several decisive margins are very small—for instance, Table 3 VinDr BI-RADS linear-probing bACC is 41.41 for GLAM versus 41.35 for MaMA and 41.26 for CLIP. In some cells GLAM is not the best (e.g., VinDr density full fine-tune bACC 78.27 versus 78.56 for Mammo-CLIP-B5). Without uncertainty quantification, the claim of consistent state-of-the-art performance is not statistically supported. Please run multiple random seeds and report mean±std or paired tests for the main comparisons.
  3. [Tab. 5 / Ablation interpretation] The ablation row labels are ambiguous and the comparison does not isolate the geometric prior. Specifically, replacing 'AP-slice sampling' with attention over all patches changes the negative set, the number of computation paths, and the effective receptive field, not just the geometric correspondence. A more controlled experiment would keep the same AP-slice grouping but permute slice indices, or use a learned/random alignment of the same complexity. As written, the ablation supports the usefulness of the local contrastive loss, but not the claim that the geometry guidance per se is the cause of the gains.
minor comments (4)
  1. [Abstract / Title] The abstract uses 'Global and Local Alignment for Multi-view mammography' while the title uses 'Geometry-Guided Local Alignment'; please align the terminology.
  2. [Table 5] The checkmark columns and the columns labeled 'M=16, M=81, M=324' are difficult to parse. Please clarify which rows correspond to which ablated component and report per-setting values in a more readable layout.
  3. [Figure 3] Please specify how query patches are selected from ROIs and how 'best match' is computed. The current caption does not state whether the same AP-slice constraint is used for the visualization or whether the match is simply the maximum-attention patch.
  4. [Sec. 3.1, Baselines] Mammo-CLIP uses official pre-trained weights while all other baselines are pre-trained on EMBED. This confounds data scale and method; the paper acknowledges it, but a brief discussion of the confound in the results section would help the reader interpret the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the geometry-guided local alignment is a self-supervised training prior, not a derivation from the downstream labels, and the reported gains are measured on held-out benchmarks.

full rationale

GLAM's central claim is an empirical pretraining result: it defines contrastive objectives (Eqs. 1–4) over unlabeled mammography image pairs and synthesized text, then evaluates the resulting encoder on held-out BI-RADS/density/cancer tasks across EMBED, VinDr, and RSNA-Mammo. The geometry-guided local alignment is an assumed imaging prior (same AP slice corresponds to related tissue), not a quantity fitted to the downstream labels; the model is trained with this prior and then tested on separate labeled data. The ablation in Table 5 shows that removing GLA hurts performance, but this is a standard ablation, not a circular redefinition. The only self-citations are using the authors' prior MaMA as a comparison baseline and borrowing its report-synthesis procedure from [10]; neither is load-bearing for the claim that geometry guidance improves alignment, and both are external, published methods rather than unverified premises. The paper's own acknowledgment of residual AP-axis misalignment is a limitation of the geometric assumption, not a circularity. No equation reduces to its input by construction, no fitted parameter is renamed as a prediction, and no uniqueness claim is imported from the authors' prior work. The evaluation is self-contained against external datasets and baselines.

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

The paper introduces no new physical or conceptual entities. Its claims rest on standard contrastive learning machinery plus domain-specific geometric assumptions about mammography views. The free parameters are standard model hyperparameters rather than fitted physical constants.

free parameters (3)
  • M (number of super-patches) = 324 (tested 16, 81, 324)
    Number of spatial regions for local alignment; chosen by hand, affects granularity of alignment and downstream performance.
  • Loss weight for L_local vs L_global = 1.0 (equal weights)
    L_final = L_global + L_local; equal weighting is an unexamined choice that could affect the balance between global and local objectives.
  • Temperature tau in contrastive loss = learnable
    Learnable temperature in Eq. (1); standard in CLIP-style losses, but its learned value is not analyzed.
assumptions (3)
  • domain assumption CC and MLO views can be aligned along the AP axis after pectoral removal and rotation, so that patches in the same AP position correspond to the same tissue in the 3D breast
    Used in Sec. 2.2 for AP slice sampling and patch-to-slice alignment. If wrong due to breast deformation, the local alignment objective may align unrelated tissue.
  • domain assumption Synthesized radiology reports from tabular data provide sufficiently informative text supervision for VLP
    Pre-processing following [10]; the quality of these synthetic reports is not directly validated.
  • domain assumption Pre-trained DiNOv2 and BioClinical-BERT are reasonable initialization encoders for mammography VLP
    Implementation choice; standard transfer-learning assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GLAM: Geometry-Guided Local Alignment for Multi-View VLP in Mammography." pith.science (2026). https://pith.science/paper/WLJLBTN6

@misc{pith2026250910344,
  author       = {Pith},
  title        = {Pith review of: GLAM: Geometry-Guided Local Alignment for Multi-View VLP in Mammography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WLJLBTN6}},
  note         = {Machine review of arXiv:2509.10344}
}
read the original abstract

Mammography screening is an essential tool for early detection of breast cancer. The speed and accuracy of mammography interpretation have the potential to be improved with deep learning methods. However, the development of a foundation visual language model (VLM) is hindered by limited data and domain differences between natural and medical images. Existing mammography VLMs, adapted from natural images, often ignore domain-specific characteristics, such as multi-view relationships in mammography. Unlike radiologists who analyze both views together to process ipsilateral correspondence, current methods treat them as independent images or do not properly model the multi-view correspondence learning, losing critical geometric context and resulting in suboptimal prediction. We propose GLAM: Global and Local Alignment for Multi-view mammography for VLM pretraining using geometry guidance. By leveraging the prior knowledge about the multi-view imaging process of mammograms, our model learns local cross-view alignments and fine-grained local features through joint global and local, visual-visual, and visual-language contrastive learning. Pretrained on EMBED [14], one of the largest open mammography datasets, our model outperforms baselines across multiple datasets under different settings.

Figures

Figures reproduced from arXiv: 2509.10344 by the authors.

Figure 1
Figure 1. The Importance of Multi-view. (a) Due to the imaging process, an ROI (red dot) appears in the same anterior-posterior (AP) slice in both mammography views. (b) Two ROIs located at different positions in the same AP slice result in the same CC view image but different MLO view images, demonstrating the single-view ambiguity. on both views to accurately locate regions of interest (ROIs), such as tumors or calcificatio… view at source ↗
Figure 2
Figure 2. GLAM Model. (a) Our method conducts global multi-view CLIP and aligns global visual features to text features from the report. (b) The patch level feature from each view is used to conduct geometry-guided local alignment, where patches from the same AP slice are used as positive matches in a cross-attention mechanism. plane during imaging. We evaluate our method on three datasets with varied distributions [14,3,23],… view at source ↗
Figure 3
Figure 3. Cross View Patch-to-Slice Attention Visualization [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Benchmarking the Robustness of Foundation Models for Mammography under Domain Shift

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Mammography-specific VLMs lead mean OOD linear-probe performance across 15 datasets, but robustness depends on pretraining objective and is highly dataset-heterogeneous, not on mammography exposure alone.

Reference graph

Works this paper leans on

35 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    Radiology292(2), 331–342 (2019)

    Akselrod-Ballin, A., et al.: Predicting breast cancer by applying deep learning to linked health records and mammograms. Radiology292(2), 331–342 (2019)

    2019

  2. [2]

    In: Proceedings of the 2nd Clinical Natural Language Processing Workshop

    Alsentzer, E., McDermott, M., et al.: Publicly available clinical BERT embeddings. In: Proceedings of the 2nd Clinical Natural Language Processing Workshop. pp. 72–78. Association for Computational Linguistics, Minneapolis, Minnesota, USA (Jun 2019). https://doi.org/10.18653/v1/W19-1909, https://www.aclweb.org/ anthology/W19-1909

    doi:10.18653/v1/w19-1909 2019

  3. [3]

    Carr, C., et.al., Y.C.: Rsna screening mammography breast cancer detection (2022), https://kaggle.com/competitions/rsna-breast-cancer-detection

    2022

  4. [4]

    In: International conference on machine learning

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597–1607. PMLR (2020)

    2020

  5. [5]

    arXiv preprint arXiv:2404.15946 (2024)

    Chen, X., Yang, X., et al.: Mammo-clip: Leveraging contrastive language-image pre- training (clip) for enhanced breast cancer diagnosis with multi-view mammography. arXiv preprint arXiv:2404.15946 (2024)

    arXiv 2024

  6. [6]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Chen, Y., Carneiro, G., et al.: Multi-view local co-occurrence and global consis- tency learning improve mammogram classification generalisation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 3–13. Springer (2022)

    2022

  7. [7]

    Medical image analysis96, 103192 (2024)

    Chen, Y., Liu, Y., Wang, C., Elliott, M., Kwok, C.F., Peña-Solorzano, C., Tian, Y., Liu, F., Frazer, H., McCarthy, D.J., et al.: Braixdet: Learning to detect malignant breast lesion with incomplete annotations. Medical image analysis96, 103192 (2024)

    2024

  8. [8]

    In: 2009 IEEE conference on CVPR

    Deng, J., Fei-Fei, L., et al.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on CVPR. pp. 248–255. Ieee (2009)

    2009

Show all 35 references
  1. [9]

    arXiv preprint arXiv:2010.11929 (2020) 10 Y

    Dosovitskiy, A., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 10 Y. Du et al

    2010 arXiv

  2. [10]

    arXiv preprint arXiv:2409.18119 (2024)

    Du, Y., Onofrey, J., Dvornek, N.C.: Multi-view and multi-scale alignment for contrastive language-image pre-training in mammography. arXiv preprint arXiv:2409.18119 (2024)

    2024 arXiv

  3. [11]

    Medical Physics 33(9), 3203–3212 (2006)

    Engeland, S.v., Timp, S., Karssemeijer, N.: Finding corresponding regions of interest in mediolateral oblique and craniocaudal mammographic views. Medical Physics 33(9), 3203–3212 (2006)

    2006

  4. [12]

    arXiv preprint arXiv:2405.12255 (2024)

    Ghosh, S., Batmanghelich, K., et al.: Mammo-clip: A vision language foundation model to enhance data efficiency and robustness in mammography. arXiv preprint arXiv:2405.12255 (2024)

    2024 arXiv

  5. [13]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Jain, K., Rangarajan, K., Arora, C.: Follow the radiologist: Clinically relevant multi-view cues for breast cancer detection from mammograms. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 102–112. Springer (2024)

    2024

  6. [14]

    Radiology: Artificial Intelligence5(1), e220047 (2023)

    Jeong, J.J., Smith, G., et al.: The emory breast imaging dataset (embed): A racially diverse, granular dataset of 3.4 million screening and diagnostic mammographic images. Radiology: Artificial Intelligence5(1), e220047 (2023)

    2023

  7. [15]

    In: MICCAI

    Ji, C., Shen, D., et al.: Mammo-net: Integrating gaze supervision and interactive information in multi-view mammogram classification. In: MICCAI. pp. 68–78. Springer (2023)

    2023

  8. [16]

    Information Fusion52, 308–321 (2019)

    Jouirou, A., Baâzaoui, A., Barhoumi, W.: Multi-view information fusion in mam- mograms: A comprehensive overview. Information Fusion52, 308–321 (2019)

    2019

  9. [17]

    In: Proceedings of the IEEE/CVF conference on CVPR

    Liu, Y., Yu, Y., et al.: Cross-view correspondence reasoning based on bipartite graph convolutional network for mammogram mass detection. In: Proceedings of the IEEE/CVF conference on CVPR. pp. 3812–3822 (2020)

    2020

  10. [18]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(10), 5947–5961 (2021)

    Liu, Y., Yu, Y., et al.: Act like a radiologist: towards reliable multi-view correspon- dence reasoning for mammogram mass detection. IEEE Transactions on Pattern Analysis and Machine Intelligence44(10), 5947–5961 (2021)

    2021

  11. [19]

    In: 2020 25th International Conference on Pattern Recognition (ICPR)

    Ma, J., Li, X., Li, H., Wang, R., Menze, B., Zheng, W.S.: Cross-view relation networks for mammogram mass detection. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 8632–8638. IEEE (2021)

    2020

  12. [20]

    Medical Image Analysis99, 103320 (2025)

    Manigrasso, F., Morra, L., et al.: Mammography classification with multi-view deep learning techniques: Investigating graph and transformer-based architectures. Medical Image Analysis99, 103320 (2025)

    2025

  13. [21]

    In: European conference on computer vision

    Mu, N., Kirillov, A., Wagner, D., Xie, S.: Slip: Self-supervision meets language-image pre-training. In: European conference on computer vision. pp. 529–544. Springer (2022)

    2022

  14. [22]

    Oquab, M., Bojanowski, P., et al.: Dinov2: Learning robust visual features without supervision (2023)

    2023

  15. [23]

    Physionet https://doi

    Pham, H.H., Trung, H.N., Nguyen, H.Q.: Vindr-mammo: A large-scale benchmark dataset for computer-aided detection and diagnosis in full-field digital mammogra- phy. Physionet https://doi. org/10.13026/br2v-7517 (2022)

    2022 doi

  16. [24]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PMLR (2021)

    2021

  17. [25]

    Medical physics33(7Part1), 2574–2585 (2006)

    Sahiner, B., Zhou, C., et al.: Joint two-view information for computerized detection of microcalcifications on mammograms. Medical physics33(7Part1), 2574–2585 (2006)

    2006

  18. [26]

    CA: a cancer journal for clinicians64(1) (2014) Geometry-Guided Local Alignment for Multi-View VLP in Mammography 11

    Siegel, R., Ma, J., Zou, Z., Jemal, A.: Cancer statistics, 2014. CA: a cancer journal for clinicians64(1) (2014) Geometry-Guided Local Alignment for Multi-View VLP in Mammography 11

    2014

  19. [27]

    In: MICCAI

    Sun, Z., Jiang, H., Ma, L., Yu, Z., Xu, H.: Transformer based multi-view network for mammographic image classification. In: MICCAI. pp. 46–54. Springer (2022)

    2022

  20. [28]

    CA: a cancer journal for clinicians71(3), 209–249 (2021)

    Sung, H., Bray, F., et al.: Global cancer statistics 2020: Globocan estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians71(3), 209–249 (2021)

    2020

  21. [29]

    Advances in Neural Information Processing Systems35, 33536–33549 (2022)

    Wang, F., Yu, L., et al.: Multi-granularity cross-modal alignment for generalized medical visual representation learning. Advances in Neural Information Processing Systems35, 33536–33549 (2022)

    2022

  22. [30]

    arXiv preprint arXiv:2210.10163 (2022)

    Wang, Z., Wu, Z., Agarwal, D., Sun, J.: Medclip: Contrastive learning from unpaired medical images and text. arXiv preprint arXiv:2210.10163 (2022)

    2022 arXiv

  23. [31]

    medRxiv pp

    Wu, C., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Medklip: Medical knowledge enhanced language-image pre-training. medRxiv pp. 2023–01 (2023)

    2023

  24. [32]

    Neurocomputing536, 21–29 (2023)

    Xia, L., Gao, Z., et al.: Neural network model based on global and local features for multi-view mammogram classification. Neurocomputing536, 21–29 (2023)

    2023

  25. [33]

    Medical Image Analysis73, 102204 (2021)

    Yang, Z., Huang, L., et al.: Momminet-v2: Mammographic multi-view mass identi- fication networks. Medical Image Analysis73, 102204 (2021)

    2021

  26. [34]

    In: International Conference on Medical Image Computing and Computer- Assisted Intervention

    You, K., Roh, B., et al.: Cxr-clip: Toward large scale chest x-ray language-image pre- training. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 101–111. Springer (2023)

    2023

  27. [35]

    In: Machine Learning for Healthcare Conference

    Zhang, Y., Jiang, H., Miura, Y., Manning, C.D., Langlotz, C.P.: Contrastive learning of medical visual representations from paired images and text. In: Machine Learning for Healthcare Conference. pp. 2–25. PMLR (2022)

    2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.