Mammo-Mamba: A Hybrid State-Space and Transformer Architecture with Sequential Mixture of Experts for Multi-View Mammography

Arash Mohammadi; Farnoush Bayatmakou; Nicole Simone; Reza Taleei

REVIEW 4 major objections 6 minor 29 references

Mammo-Mamba: A Hybrid State-Space and Transformer Architecture with Sequential Mixture of Experts for Multi-View Mammography

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

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

Pith's one-line read Mammo-Mamba, a hybrid state-space and transformer architecture with sequential expert gating, reports top results on the CBIS-DDSM multi-view mammogram benchmark.

desk verdict A competent incremental architecture paper whose 'superior performance' claim is undercut by a breast-level split that risks patient leakage and an unmeasured efficiency claim. read the letter →

arxiv 2507.17662 v1 pith:XPO7CO3M submitted 2025-07-23 eess.IV cs.CVcs.LG

Farnoush Bayatmakou , Reza Taleei , Nicole Simone , Arash Mohammadi This is my paper · ORCID

classification eess.IVcs.CVcs.LG

keywords multi-viewmammographybreastcancerclassificationstate-spacemodelsMambamixtureofexpertstransformerCBIS-DDSMcomputer-aideddiagnosis

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 sets out to establish that a dual-stream architecture pairing the MambaVision backbone with a new Sequential Mixture of Experts (SeqMoE) mechanism can classify multi-view mammograms more accurately than transformer-only or Mamba-only models. The authors argue that standard transformer attention scales quadratically with the number of image patches, whereas selective state-space models offer linear-time processing, making them a promising backbone for high-resolution mammography. Their SecMamba block inserts a binary gate after each deep block, letting the model choose per position whether to apply one transformation or a stacked deeper transformation, which they call expert routing along the depth dimension. On the CBIS-DDSM mass classification benchmark, Mammo-Mamba is reported to reach 0.8696 accuracy and 0.9089 AUC, outperforming the listed baselines and state-of-the-art methods.

What carries the argument

The central object is the SecMamba block, a modified MambaVision block whose state-space output is selectively modulated by a sigmoid gate and then passed through the Sequential Mixture of Experts (SeqMoE) gate. SeqMoE is a pairwise depth mixture of experts: rather than mixing parallel experts, it pairs each block with its predecessor and learns whether to apply one transformation or a second stacked transformation, with a softmax gate over pooled representations. This gated interpolation acts as an adaptive residual connection and as dynamic depth routing. The architecture also uses dual-stream fusion: cropped images and whole mammograms each go through their own copy of the modified stages, and the resulting embeddings are concatenated and gated before a multilayer perceptron classifier.

What would settle it

Retrain Mammo-Mamba and the same baselines on CBIS-DDSM using a patient-level split, where all images of a patient stay in one partition, and compare accuracy and AUC; if the margins shrink or flip, the claim of superior classification performance does not generalize to leakage-free evaluation.

Watch

Extended reading notes

Core claim

The central claim is that a hierarchical vision backbone combining convolutional layers, selective state-space models, and transformer attention can be improved for multi-view mammography by inserting a Sequential Mixture of Experts gate after each deep block, and that this combination outperforms both pure transformer and pure Mamba baselines. The SecMamba block treats each block as an expert in a chain; the gate takes the block input and output, mean-pools them, and emits a scalar that interpolates between them, so the effective depth of processing varies with the content of each token. Placed in stages 3 and 4 of MambaVision, with separate streams for cropped lesion regions and whole mammograms fused by a gated combination, the model yields the reported accuracy of 0.8696 and AUC of 0.9089. The authors further claim the design preserves linear-time complexity, avoiding the quadratic patch scaling of standard Transformers.

Load-bearing premise

The reported gains assume that the breast-level split keeps every patient's data out of the opposite set; if the same patient's left and right breasts land in different splits, the test set is no longer independent and the margins may be inflated.

Editorial extensions

If this is right

If the reported results hold, multi-view mammogram classification can be done with a linear-time state-space backbone instead of quadratic-complexity transformer-only models, making high-resolution inputs more feasible.
The SeqMoE gate gives the model a content-adaptive choice of processing depth at each position, which the authors connect to dynamic unrolling and early-exit style routing.
The dual-stream design (cropped lesion plus whole mammogram) is what the ablation attributes the largest single accuracy gain to, moving from 0.7517 to 0.8322 when added to the backbone.
Removing SeqMoE from the full model drops accuracy from 0.8792 to 0.8322, indicating the gate contributes beyond the dual-stream fusion alone.
The architecture provides a template for further state-space based medical image classifiers, and the paper cites its adaptation of MambaVision to multi-view mammography as one of the first of its kind.

Reading between the lines

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

Because the dataset split is performed at the breast level, a patient with two breasts in the set could contribute one breast to training and the other to testing; a patient-level split is a natural retest that would clarify whether the margin over baselines persists.
The SeqMoE gate is architecture-agnostic: the same pairwise depth gating could be added to other hierarchical vision backbones, and a reader could test whether it helps on other lesion-imaging tasks with small datasets.
The gate's learned values resemble a dynamic difficulty signal, and a natural extension would use them to skip expensive blocks at inference time, something the paper does not report.
A direct comparison of training and inference throughput between Mammo-Mamba and transformer-only baselines would quantify the claimed computational-efficiency advantage, since the paper's efficiency argument rests on the state-space formulation rather than on measured runtime.

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.

Desk Editor's Note

read the letter

The headline: a competent, incremental architecture paper whose central 'superior performance' claim is currently undercut by a patient-leakage risk in the split and an unmeasured efficiency claim. The MambaVision backbone for multi-view mammography is a sensible idea, and the dual-stream crop+whole design with the iterative scalar depth gate is new as a combination. The ablations do show the gate and dual-stream each add something, and the final model was run seven times with standard deviations, which is more than many papers in this space do.

The soft spots are real. First, the split. The paper says 'Dataset splitting was performed at the breast level to avoid patient-level data leakage.' That is wrong. CBIS-DDSM contains multiple exams per patient; splitting by breast can place the same patient's two breasts into train and test. Contralateral breasts share density and tissue texture, so the model can exploit identity rather than lesion pathology. This inflates all absolute numbers, and while it likely inflates baselines too, the paper's claim of 'superior classification performance' over state-of-the-art is not trustworthy without a patient-disjoint split reported. This is the main thing to fix.

Second, the abstract promises 'computational efficiency' but no FLOPs or latency are reported anywhere. That's an unsupported half of the central claim.

Third, Table IV reports the best accuracy 0.8792/AUC 0.9249, while Table III reports the mean over 7 runs, 0.8696/0.9089. That's acceptable practice if labeled, but here it looks like cherry-picking; the ablation should report the same average metric as the main table.

Fourth, calling the scalar residual gate a 'Sequential Mixture of Experts' is a stretch. Eq. 13 is a learned interpolation between input and output, conceptually a highway/GRU gate without recurrence. That reduces novelty but doesn't change the empirical question.

The citation pattern is fine. Prior work cited is relevant. No code or data release, which matters for reproducibility.

Bottom line: this deserves a serious referee, but only with a request for major revision: patient-level split with multiple seeds, confidence intervals for all baselines, actual efficiency numbers, and a code release. If those come, the paper could be a useful addition to the medical imaging literature. As it stands, it's a solid engineering report with an inflated abstract.

Referee Report

4 major / 6 minor

Summary. The paper proposes Mammo-Mamba, a dual-stream architecture for multi-view mammography classification built on the MambaVision backbone. The method introduces SecMamba blocks that combine selective state-space models with self-attention, and a Sequential Mixture of Experts (SeqMoE) gate that blends each block's input and output through a learned scalar weight. The two streams process cropped lesion images and whole mammograms, and their embeddings are fused for classification. Evaluated on a mass subset of CBIS-DDSM (653 examinations, split into 504 training and 149 testing), the paper reports accuracy 0.8696 +/- 0.0109 and AUC 0.9089 +/- 0.0135 over 7 runs, claiming superior performance over Mamba-based baselines and prior state-of-the-art models, with an ablation that separates the contributions of the dual-stream design and SeqMoE.

Significance. If the empirical claims hold, the paper would offer a linear-complexity hybrid alternative to transformer-only multi-view mammography models, and the SeqMoE depth-gating idea could be reusable in other Mamba-based medical vision backbones. The manuscript has several strengths: it compares against multiple Mamba variants and mammography-specific models, includes a component-level ablation, and reports mean and standard deviation for the proposed model over multiple runs. However, the experimental protocol currently carries a patient-leakage risk, the computational-efficiency claim is unmeasured, and the 'superior across all key metrics' conclusion lacks statistical support because the in-house baselines are reported without variance or significance tests. The central architecture claim is also underevidenced: SeqMoE is a scalar gated residual connection, and no mechanism-level analysis is provided to show that expert-style routing occurs.

major comments (4)

[§IV, Dataset] The sentence 'Dataset splitting was performed at the breast level to avoid patient-level data leakage' is internally contradictory. CBIS-DDSM contains multiple examinations per patient, including both breasts; a random breast-level split can place the same patient's contralateral breast in the training set and the other in the test set. Because contralateral mammograms share tissue density and anatomical structure, the model can exploit patient identity rather than lesion pathology, which would inflate all reported margins in Tables I–IV. Please re-split at the patient level (all breasts and examinations of a patient in the same partition) and re-run the complete comparison before the performance claims can be accepted.
[Tables I, II, IV] The in-house baselines are reported as single point estimates without confidence intervals, and no significance tests are provided. The headline claim of 'superior classification performance across all key metrics' is not statistically supported; for example, the 4.4-point accuracy gap over MambaVision-L in Table I may be within run-to-run variability. Additionally, Table II reports Mammo-Mamba accuracy 0.8792 and AUC 0.9249, while Table III reports 0.8696 +/- 0.0109 and 0.9089 +/- 0.0135; it is unclear whether Table II is the best run and Table III the mean. Please report mean +/- standard deviation for every in-house model and include pairwise significance tests or interval-overlap analysis for the key comparisons.
[Abstract and §IV] The claim that Mammo-Mamba achieves 'superior classification performance across all key metrics while maintaining computational efficiency' is unsupported with respect to efficiency. No FLOPs, latency, parameter count, GPU memory, or throughput measurement is reported anywhere in Section IV. Either add efficiency measurements for Mammo-Mamba and the Mamba/transformer baselines, or remove the efficiency claim from the abstract and introduction.
[§III.B, Eqs. (12)–(13)] The SeqMoE mechanism defined by Eqs. (12)–(13) is a scalar gated residual connection between a block's input and output, not a mixture of experts in the conventional sense; there is no set of parallel expert modules, no expert-selection objective, and no load-balancing mechanism. The ablation in Table IV shows only that the full model beats the dual-stream-only model, and it provides no probe of gate behavior across classes, views, or stages. This leaves the paper's central architectural contribution—'expert-driven feature refinement'—underevidenced and the terminology potentially misleading. Please provide mechanism-level analyses (e.g., gate value distributions, expert utilization across inputs) or revise the framing to 'adaptive depth gating.'

minor comments (6)

[§II.A] The description 'We used the train-test split recommendation for the dataset' is vague; if the 504/149 split is a random split, please state the random seed and splitting procedure, and if it is derived from the CBIS-DDSM official split, please explain how the 653 selected examinations map to that partition.
[Table I] The MambaVision T2 row reports F1 = 0.8436 with accuracy 0.7778, sensitivity 0.7175, and specificity 0.8196; such a high F1 with lower sensitivity and specificity appears internally inconsistent, so please verify the reported values.
[§III.B] Equation (12) has a formatting break ('W G out (12) ·ReLU...') and uses G for the SeqMoE gate while Eq. (9) already uses G for the selective modulation gate; please use distinct symbols (e.g., g_seq and g_mod) throughout Section III to avoid confusion.
[Abstract and Table III] The 'superior across all key metrics' statement is not fully supported because several state-of-the-art rows in Table III have missing sensitivity, specificity, or F1 entries; please clarify which metrics are compared on the same patient/exam basis and which comparisons are complete.
[§I and §II.B] The paper states that MambaVision provides linear scaling and lower GPU memory usage but gives no complexity analysis or runtime evidence in the present manuscript; if these properties are invoked to support efficiency, please cite the original MambaVision analysis and, ideally, measure them in the experimental section.
[Figure 2] The caption says 'iterative gating after each expert,' but the gating in Eq. (13) is applied between each block and its input; please align the caption terminology with the equations so readers can match the figure to the formal definition.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Mammo-Mamba's performance claim is an empirical benchmark result against an external dataset; the architecture builds on externally published MambaVision, and no fitted parameter is repackaged as a prediction.

full rationale

The paper contains no analytic derivation that reduces to its inputs. The central claim is empirical: Mammo-Mamba achieves reported accuracy/AUC on CBIS-DDSM, an external benchmark. The architecture is constructed from MambaVision [21] with added SecMamba blocks and a SeqMoE gate defined by explicit equations (Eqs. 1-13); these components are not defined in terms of the target labels or the reported metrics. The baselines in Tables I-III are independent models, and the only author self-citation [12] is used as a comparison baseline, not as load-bearing justification for the method. The statement 'Dataset splitting was performed at the breast level to avoid patient-level data leakage' is a potential experimental-protocol weakness (breast-level splitting does not guarantee patient-disjoint partitions), and the 'computational efficiency' claim is not accompanied by FLOPs or latency measurements; both are correctness/validity concerns, not circularity. No fitted parameter, gate, or architectural choice is renamed as a prediction, and no external result is imported from the authors' prior work to force the outcome. The derivation chain is therefore self-contained and non-circular.

Assumptions & free parameters 2 free parameters · 3 assumptions · 1 invented entities

The empirical claims rest on an unverified data-splitting assumption, unreported exact hyperparameters, transfer of pretrained weights, and the assumed efficiency of the Mamba implementation. The invented SeqMoE gate is only validated through end-task accuracy.

free parameters (2)

Training hyperparameters (learning rate and weight decay) = lr 5e-5 to 7e-5, weight decay 1e-3 to 3e-3, exact values not disclosed
The paper reports ranges instead of the values used for each model, so final comparisons depend on unstated tuning choices that could favor the proposed model.
Dataset split (504 training / 149 testing exams) = 504/149 from 653 selected mass cases
The superior performance claim is based on a single split with no repeated cross-validation for baselines; the split choice is a hand-selected experimental condition that directly shapes the headline numbers.

assumptions (3)

domain assumption The MambaVision backbone's pretrained weights transfer to grayscale mammography at 224x224 resolution.
Section III-C states the patch embedding is initialized with MambaVision values and stages 3 and 4 are fine-tuned; no domain adaptation or self-supervised pretraining is used.
domain assumption The breast-level split prevents patient-level leakage.
Section IV says splitting at breast level avoids patient-level leakage, but patient-level deduplication is not described; if a patient has two breasts in the dataset, one could be in train and one in test.
domain assumption The standard selective-scan convolution in Eqs. (6) and (7) delivers the claimed linear-time efficiency.
Efficiency claims rely on the Mamba implementation's complexity; no FLOPs, memory, or latency measurements are provided in the paper.

invented entities (1)

Sequential Mixture of Experts (SeqMoE) depth gate
purpose: Content-adaptive scalar interpolation between a SecMamba block's output and the previous block's output (Eq. 13), presented as the main architectural contribution.
The mechanism is a learned residual gate, not sparse expert routing; no load balancing, no expert-specific losses, and no isolated behavioral evidence are provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mammo-Mamba: A Hybrid State-Space and Transformer Architecture with Sequential Mixture of Experts for Multi-View Mammography." pith.science (2026). https://pith.science/paper/XPO7CO3M

@misc{pith2026250717662,
  author       = {Pith},
  title        = {Pith review of: Mammo-Mamba: A Hybrid State-Space and Transformer Architecture with Sequential Mixture of Experts for Multi-View Mammography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XPO7CO3M}},
  note         = {Machine review of arXiv:2507.17662}
}

read the original abstract

Breast cancer (BC) remains one of the leading causes of cancer-related mortality among women, despite recent advances in Computer-Aided Diagnosis (CAD) systems. Accurate and efficient interpretation of multi-view mammograms is essential for early detection, driving a surge of interest in Artificial Intelligence (AI)-powered CAD models. While state-of-the-art multi-view mammogram classification models are largely based on Transformer architectures, their computational complexity scales quadratically with the number of image patches, highlighting the need for more efficient alternatives. To address this challenge, we propose Mammo-Mamba, a novel framework that integrates Selective State-Space Models (SSMs), transformer-based attention, and expert-driven feature refinement into a unified architecture. Mammo-Mamba extends the MambaVision backbone by introducing the Sequential Mixture of Experts (SeqMoE) mechanism through its customized SecMamba block. The SecMamba is a modified MambaVision block that enhances representation learning in high-resolution mammographic images by enabling content-adaptive feature refinement. These blocks are integrated into the deeper stages of MambaVision, allowing the model to progressively adjust feature emphasis through dynamic expert gating, effectively mitigating the limitations of traditional Transformer models. Evaluated on the CBIS-DDSM benchmark dataset, Mammo-Mamba achieves superior classification performance across all key metrics while maintaining computational efficiency.

Figures

Figures reproduced from arXiv: 2507.17662 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 22 canonical work pages

[1]

Global Cancer Statistics 2020: Globocan Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries,

H. Sung, et al., “Global Cancer Statistics 2020: Globocan Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries,” CA Cancer. J. Clin. , vol. 71, no. 3, pp. 209–249, 2021

work page 2020
[2]

Novel use of Scout® Reflector for Target Localization in Preoperative Breast Radiotherapy,

N. Ghassemi, et al. , “Novel use of Scout® Reflector for Target Localization in Preoperative Breast Radiotherapy,” Int. J. Radiat. Oncol. Biol. Phys. , vol. 120, no. 2, p. e307, 2024

work page 2024
[3]

Po91: Evaluation of Advanced Collapsed Cone Engine (ACE) Treatment Planning Algorithm for HDR Breast Plans,

S. Rooks, R. Taleei, N. Simone, P. R. Anne, and F. Mourtada, “Po91: Evaluation of Advanced Collapsed Cone Engine (ACE) Treatment Planning Algorithm for HDR Breast Plans,” Brachytherapy, vol. 22, no. 5, 2023

work page 2023
[4]

A Novel Method for Measuring the Burden of Breast Cancer in Neighborhoods,

R. K. McIntire, et al., “A Novel Method for Measuring the Burden of Breast Cancer in Neighborhoods,” Prev. Med. Rep., vol. 33, p. 102218, 2023

work page 2023
[5]

Diagnostic Performance of Digital versus Film Mammography for Breast-Cancer Screening,

E. D. Pisano, et al., “Diagnostic Performance of Digital versus Film Mammography for Breast-Cancer Screening,” N. Engl. J. Med. , vol. 353, no. 17, pp. 1773–1783, 2005

work page 2005
[6]

A review of computer aided detection in mammography,

J. Katzen and K. Dodelzon, “A review of computer aided detection in mammography,” Clinical imaging, vol. 52, pp. 305–309, 2018

work page 2018
[7]

Multi-view convolu- tional neural networks for mammographic image classification,

L. Sun, J. Wang, Z. Hu, Y . Xu, and Z. Cui, “Multi-view convolu- tional neural networks for mammographic image classification,” IEEE Access, vol. 7, pp. 126 273–126 282, 2019

work page 2019
[8]

Multi-view feature fusion based four views model for mammo- gram classification using convolutional neural network,

H. N. Khan, A. R. Shahid, B. Raza, A. H. Dar, and H. Alquhayz, “Multi-view feature fusion based four views model for mammo- gram classification using convolutional neural network,” IEEE Access, vol. 7, pp. 165 724–165 733, 2019

work page 2019

Show all 29 references

[9]

Deep learning to improve breast cancer detection on screening mammography,

L. Shen, L. R. Margolies, J. H. Rothstein, E. Fluder, R. McBride, and W. Sieh, “Deep learning to improve breast cancer detection on screening mammography,” Sci. Rep., vol. 9, no. 1, p. 12495, 2019

2019
[10]

An in- tegrated framework for breast mass classification and diagnosis using stacked ensemble of residual neural networks,

A. Baccouche, B. Garcia-Zapirain, and A. S. Elmaghraby, “An in- tegrated framework for breast mass classification and diagnosis using stacked ensemble of residual neural networks,”Sci. Rep., vol. 12, no. 1, p. 12259, 2022

2022
[11]

Mv-swin-t: mam- mogram classification with multi-view swin transformer,

S. Sarker, P. Sarker, G. Bebis, and A. Tavakkoli, “Mv-swin-t: mam- mogram classification with multi-view swin transformer,” in IEEE International Symposium on Biomedical Imaging , 2024, pp. 1–5

2024
[12]

Integrating ai for human-centric breast cancer diagnostics: A multi- scale and multi-view swin transformer framework,

F. Bayatmakou, R. Taleei, M. A. Toutounchian, and A. Mohammadi, “Integrating ai for human-centric breast cancer diagnostics: A multi- scale and multi-view swin transformer framework,” in IEEE Interna- tional Conference on Human-Machine Systems , 2025

2025
[13]

Advancing precision in breast can- cer detection: a fusion of vision transformers and cnns for calcification mammography classification,

S. S. Boudouh and M. Bouakkaz, “Advancing precision in breast can- cer detection: a fusion of vision transformers and cnns for calcification mammography classification,” Appl. Intell., vol. 54, no. 17, pp. 8170– 8183, 2024

2024
[14]

Hybridmammonet: A hybrid cnn-vit architecture for multi-view mammography image classifica- tion,

H. Allaoui, Y . Alj, and Y . Ameskine, “Hybridmammonet: A hybrid cnn-vit architecture for multi-view mammography image classifica- tion,” in IEEE 12th International Symposium on Signal, Image, Video and Communications, 2024, pp. 1–6

2024
[15]

Mamba: Linear-time sequence modeling with selective state spaces,

A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

2023 arXiv
[16]

Vision mamba: Efficient visual representation learning with bidirectional state space model,

L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,” arXiv preprint arXiv:2401.09417 , 2024

2024 arXiv
[17]

Vmamba: Visual state space model,

Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “Vmamba: Visual state space model,” Advances in neural information processing systems , vol. 37, pp. 103 031–103 063, 2024

2024
[18]

Multi-scale vmamba: Hierarchy in hierarchy visual state space model,

Y . Shi, M. Dong, and C. Xu, “Multi-scale vmamba: Hierarchy in hierarchy visual state space model,” Advances in Neural Information Processing Systems, vol. 37, pp. 25 687–25 708, 2024

2024
[19]

Medmamba: Vision mamba for medical image classification,

Y . Yue and Z. Li, “Medmamba: Vision mamba for medical image classification,” arXiv preprint arXiv:2403.03849 , 2024

2024 arXiv
[20]

Dynamic vision mamba,

M. Wu, Z. Li, Z. Liang, M. Li, X. Zhao, S. Khaki, Z. Zhu, X. Peng, K. N. Plataniotis, K. Wang et al. , “Dynamic vision mamba,” arXiv preprint arXiv:2504.04787, 2025

2025 arXiv
[21]

Mambavision: A hybrid mamba- transformer vision backbone,

A. Hatamizadeh and J. Kautz, “Mambavision: A hybrid mamba- transformer vision backbone,” in Proceedings of the Computer Vision and Pattern Recognition Conference , 2025, pp. 25 261–25 270

2025
[22]

Lora: Low-rank adaptation of large language models

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al., “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

2022
[23]

A curated mammography data set for use in computer-aided detection and diagnosis research,

R. S. Lee, et al. , “A curated mammography data set for use in computer-aided detection and diagnosis research,” Scientific data , vol. 4, no. 1, pp. 1–9, 2017

2017
[24]

Can score-based generative modeling effectively handle medical image classification?

S. Sarker, P. Sarker, G. Bebis, and A. Tavakkoli, “Can score-based generative modeling effectively handle medical image classification?” arXiv preprint arXiv:2502.17727 , 2025

2025 arXiv
[25]

Exploiting patch sizes and resolutions for multi-scale deep learning in mammogram image classification,

G. I. Quintana, Z. Li, L. Vancamberg, M. Mougeot, A. Desolneux, and S. Muller, “Exploiting patch sizes and resolutions for multi-scale deep learning in mammogram image classification,” Bioengineering, vol. 10, no. 5, p. 534, 2023

2023
[26]

Deep cross-training: An approach to improve deep neural network classification on mam- mographic images,

K. L. dos Santos and M. P. dos Santos Silva, “Deep cross-training: An approach to improve deep neural network classification on mam- mographic images,” Expert Syst. Appl. , vol. 238, p. 122142, 2024

2024
[27]

Mammo-clustering: A weakly supervised multi-view global-local context clustering network for detection and classification in mammography,

S. Yang, et al., “Mammo-clustering: A weakly supervised multi-view global-local context clustering network for detection and classification in mammography,” arXiv preprint arXiv:2409.14876 , 2024

2024 arXiv
[28]

An open codebase for enhancing transparency in deep learning-based breast cancer diagnosis utilizing cbis-ddsm data,

L. Liao and E. M. Aagaard, “An open codebase for enhancing transparency in deep learning-based breast cancer diagnosis utilizing cbis-ddsm data,” Sci. Rep., vol. 14, no. 1, p. 27318, 2024

2024
[29]

Xfmamba: Cross-fusion mamba for multi-view medical image classification,

X. Zheng, X. Chen, S. Gong, X. Griffin, and G. Slabaugh, “Xfmamba: Cross-fusion mamba for multi-view medical image classification,” arXiv preprint arXiv:2503.02619 , 2025

2025 arXiv

Pith tools

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

[1] [1]

Global Cancer Statistics 2020: Globocan Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries,

H. Sung, et al., “Global Cancer Statistics 2020: Globocan Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries,” CA Cancer. J. Clin. , vol. 71, no. 3, pp. 209–249, 2021

work page 2020

[2] [2]

Novel use of Scout® Reflector for Target Localization in Preoperative Breast Radiotherapy,

N. Ghassemi, et al. , “Novel use of Scout® Reflector for Target Localization in Preoperative Breast Radiotherapy,” Int. J. Radiat. Oncol. Biol. Phys. , vol. 120, no. 2, p. e307, 2024

work page 2024

[3] [3]

Po91: Evaluation of Advanced Collapsed Cone Engine (ACE) Treatment Planning Algorithm for HDR Breast Plans,

S. Rooks, R. Taleei, N. Simone, P. R. Anne, and F. Mourtada, “Po91: Evaluation of Advanced Collapsed Cone Engine (ACE) Treatment Planning Algorithm for HDR Breast Plans,” Brachytherapy, vol. 22, no. 5, 2023

work page 2023

[4] [4]

A Novel Method for Measuring the Burden of Breast Cancer in Neighborhoods,

R. K. McIntire, et al., “A Novel Method for Measuring the Burden of Breast Cancer in Neighborhoods,” Prev. Med. Rep., vol. 33, p. 102218, 2023

work page 2023

[5] [5]

Diagnostic Performance of Digital versus Film Mammography for Breast-Cancer Screening,

E. D. Pisano, et al., “Diagnostic Performance of Digital versus Film Mammography for Breast-Cancer Screening,” N. Engl. J. Med. , vol. 353, no. 17, pp. 1773–1783, 2005

work page 2005

[6] [6]

A review of computer aided detection in mammography,

J. Katzen and K. Dodelzon, “A review of computer aided detection in mammography,” Clinical imaging, vol. 52, pp. 305–309, 2018

work page 2018

[7] [7]

Multi-view convolu- tional neural networks for mammographic image classification,

L. Sun, J. Wang, Z. Hu, Y . Xu, and Z. Cui, “Multi-view convolu- tional neural networks for mammographic image classification,” IEEE Access, vol. 7, pp. 126 273–126 282, 2019

work page 2019

[8] [8]

Multi-view feature fusion based four views model for mammo- gram classification using convolutional neural network,

H. N. Khan, A. R. Shahid, B. Raza, A. H. Dar, and H. Alquhayz, “Multi-view feature fusion based four views model for mammo- gram classification using convolutional neural network,” IEEE Access, vol. 7, pp. 165 724–165 733, 2019

work page 2019