Pith. sign in

REVIEW 4 major objections 4 minor 32 references

Fusion tokens let frozen AI see both mammogram views at once

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-08 10:08 UTC pith:JMT3MXC6

load-bearing objection A reasonable architectural contribution for multi-view mammography with token-based fusion, but the evidence for 'consistent improvements' is thin without error bars and with per-dataset hyperparameter tuning on small validation sets. the 4 major comments →

arxiv 2607.06309 v1 pith:JMT3MXC6 submitted 2026-07-07 cs.CV cs.AI

Token-Based Dual-view Fusion and Adaptation of Large Vision Models for Breast Cancer Classification

Aysan Ghayouri Pirsoltan , Shima Babakordi , Mohammad Reza Mohammadi This is my paper
classification cs.CV cs.AI
keywords multi-view fusionvision transformerprompt learningbreast cancer classificationmammographytoken-based fusionparameter-efficient transfer learningcross-attention
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes a two-stage method for classifying mammograms by combining two standard breast-imaging views (craniocaudal and mediolateral oblique) inside a large pretrained vision transformer whose backbone weights never change. In the first stage, shared learnable prompt tokens are injected into the transformer to adapt both views into a common representation space. In the second stage, dedicated fusion tokens are created via bidirectional cross-attention between the two views at selected transformer layers: each view summarizes what it sees in the other view into a single compact token, which is then prepended to that view's token sequence and processed by subsequent layers. This means cross-view information persists as a distinct, refinable signal through the network rather than being mixed into view-specific features via residual addition or concatenation. The authors argue this token-centric design avoids entangling view-specific and cross-view representations and enables progressive interaction across multiple depths of the encoder. On VinDr-Mammo and CMMD datasets, the full framework outperforms linear probing, prompt-only adaptation, residual cross-attention fusion, and a feature-level dual-view baseline (DIVF), achieving 50.40% F1 and 0.8090 AUC on the 5-class VinDr-Mammo task and a 0.10 AUC improvement over DIVF in the binary setting.

Core claim

The central object is the fusion token: a single learnable token per view, produced by mean-pooling the output of bidirectional cross-attention between CC and MLO hidden states, that is inserted into the transformer's token sequence at chosen layers and then refined by subsequent frozen layers. The paper shows that this mechanism—cross-view information carried as discrete tokens rather than merged into feature maps—consistently outperforms residual cross-attention fusion under identical training conditions, with gains of 5.22 F1 percentage points and 0.0521 AUC on the 5-class VinDr-Mammo task. The authors also find that two well-spaced fusion blocks (e.g., at layers 12 and 23 of a 27-layer-0

What carries the argument

Frozen MedSigLIP vision transformer + deep shared prompt tokens (Stage 1) + cross-view fusion blocks with bidirectional cross-attention producing mean-pooled fusion tokens inserted at multiple transformer depths (Stage 2)

Load-bearing premise

The number and placement of fusion blocks are tuned separately for each dataset using the validation set, and the paper does not test whether these hyperparameters transfer across datasets. With only two fusion blocks and dataset-specific tuning, the chosen configuration may partly reflect validation-set noise rather than a general architectural principle.

What would settle it

If the fusion-token mechanism's advantage disappears when fusion-block placement is fixed across datasets rather than tuned per dataset, the gains would be attributable to hyperparameter search rather than the token-based design itself.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Token-based fusion could generalize beyond mammography to any multi-view medical imaging task (e.g., dual-plane MRI, multi-angulation CT) where views carry complementary anatomical information and the backbone is frozen.
  • The finding that two well-spaced fusion blocks outperform both a single block and four blocks suggests there may be principled layer-spacing heuristics for where cross-view interaction is most informative in transformer encoders.
  • If the fusion-token mechanism is robust, it could replace residual cross-attention in existing multi-view transformer pipelines without unfreezing the backbone, reducing both compute and the risk of catastrophic forgetting.
  • The shared-prompt-then-fuse training order implies a curriculum: first align views in a common space, then introduce cross-view interaction. This two-stage recipe may apply to other multi-modal or multi-input adaptation problems with frozen backbones.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The optimal fusion-block placement (layers 12 and 23 for binary VinDr-Mammo) is tuned per dataset on the validation set. If this placement reflects a general principle—e.g., interaction is most useful at roughly 40% and 85% of encoder depth—it would transfer across datasets without per-dataset tuning. The paper does not test this.
  • The fusion-token design is architecturally similar to register tokens or memory tokens in vision transformers. The connection suggests that cross-view fusion tokens may function as a lightweight cross-view memory bank, and that attention-based pooling of these tokens (which underperformed mean pooling here) might benefit from a larger token count or different initialization.
  • The two-stage training (prompts first, then fusion) resembles a curriculum where alignment precedes interaction. If verified, this ordering principle could inform multi-modal foundation-model adaptation more broadly.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a two-stage framework for dual-view mammogram classification using a frozen vision transformer backbone (MedSigLIP). Stage 1 introduces deep shared prompt tokens across CC and MLO views for parameter-efficient adaptation. Stage 2 inserts dedicated cross-view fusion tokens via bidirectional cross-attention at multiple transformer depths, allowing progressive interaction across the encoder hierarchy. The framework is evaluated on VinDr-Mammo (binary and 5-class) and CMMD datasets, showing improvements over linear probing, prompt-only, and single-stage fusion baselines. The code is publicly available.

Significance. The token-based multi-depth fusion mechanism is a reasonable architectural contribution to multi-view medical image analysis, offering a structured alternative to residual cross-attention. The provision of publicly available code is a notable strength for reproducibility. The ablation studies isolating the fusion strategy, aggregation method, and prompt depth provide useful empirical insights for adapting frozen vision foundation models to multi-view settings.

major comments (4)
  1. Table 2 and the claim of 'consistent improvements': The full framework (64.96% F1) underperforms the fusion-only ablation (65.20% F1) on CMMD. Combined with the absence of error bars, confidence intervals, or significance tests across all experiments, the central claim of consistent improvement is not statistically established. Given that fusion block placement and prompt depth are tuned per-dataset on validation sets (§5.1), the reported differences could fall within run-to-run variance. The authors should either soften the 'consistent' claim or provide multi-seed results to verify that the gains are statistically significant.
  2. Reference [26] is cited as 'MedSigLIP' throughout the text (e.g., §3, §5.1), but the bibliography entry is titled 'Medgemma technical report' (arXiv:2507.05201). MedGemma and MedSigLIP are distinct models. This mismatch must be resolved to ensure reproducibility, as the backbone's pretrained capacity fundamentally affects the interpretation of the baseline comparisons.
  3. §5.1: The number and placement of fusion modules and prompt depth are tuned separately for each dataset using the validation set. With only two fusion blocks selected for VinDr-Mammo (e.g., layers 12 and 23 per §5.5), the risk of fitting to validation-set noise is non-trivial. The paper does not test cross-dataset transfer of these hyperparameters. The authors should discuss this overfitting risk and ideally report the sensitivity of the results to these specific architectural choices.
  4. §5.4.2 and Table 3: The comparison with existing methods is limited to DIVF [20], which is not a recent state-of-the-art method. The related work section discusses more recent and relevant multi-view transformer approaches (e.g., MaMVT [17], TransCorNet [24], BRAIxMVCCL [5]). The paper would be substantially strengthened by comparing against at least one of these recent baselines under identical backbone settings to properly contextualize the proposed framework's contributions.
minor comments (4)
  1. Table 1: The sample counts for VinDr-Mammo are reported using BI-RADS categories, but the text in §4 states that for the binary setting, BI-RADS 2 was grouped as Suspicious Benign and BI-RADS 4 and 5 as Suspicious Malignancy. It is unclear what happens to BI-RADS 1 and 3 in the binary setting. Clarification on the exact class definitions for the binary task would help readers.
  2. Figure 3b: The legend states 'N = 4 (AUC=0.8309)', but the text in §5.5 and Table 3 report the 4-block configuration as AUC=0.8313. Please correct this discrepancy.
  3. §3, Stage 1: The notation for the prompt set uses a tilde over p (̃ p_i), but the equation H_i = V_i([̃ p_i, H_{i-1}]) does not explicitly define how the prompt token is concatenated (e.g., prepended vs. appended). A brief clarification would improve readability.
  4. Table 2: The 'Stage 2: Cross-View Fusion Only' row is bolded for CMMD F1-Score, but the text in §5.4.1 states the full framework achieves the best overall trade-off. Ensure the bolding convention (best result per metric) is applied consistently and explained in the caption.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the careful and constructive review. The referee raises four major points: (1) the absence of statistical significance testing and the claim of 'consistent improvements' given that the full framework underperforms the fusion-only ablation on CMMD F1; (2) a bibliographic mismatch where reference [26] is cited as MedSigLIP but the entry points to the MedGemma technical report; (3) the risk of overfitting architectural hyperparameters to validation-set noise without cross-dataset transfer experiments; and (4) the limited comparison to only DIVF [20], omitting more recent multi-view transformer baselines. We address each point below and commit to revisions for all four.

read point-by-point responses

  1. Referee: Table 2 and the claim of 'consistent improvements': The full framework (64.96% F1) underperforms the fusion-only ablation (65.20% F1) on CMMD. Combined with the absence of error bars, confidence intervals, or significance tests across all experiments, the central claim of consistent improvement is not statistically established. Given that fusion block placement and prompt depth are tuned per-dataset on validation sets (§5.1), the reported differences could fall within run-to-run variance. The authors should either soften the 'consistent' claim or provide multi-seed results to verify that the gains are statistically significant.

    Authors: The referee is correct on both counts. First, the full framework does underperform the fusion-only variant on CMMD F1 (64.96% vs. 65.20%), so the word 'consistent' as applied to every metric in every setting is not fully supported by the data. Second, without multi-seed runs or significance tests, we cannot rule out that the reported differences fall within run-to-run variance. We will address this in the revision by: (a) running each main experiment with at least three random seeds and reporting mean ± standard deviation; (b) adding paired significance tests (e.g., bootstrap or DeLong's test for AUC) where appropriate; and (c) revising the language from 'consistent improvements' to a more precise statement — namely, that the full framework achieves the best AUC on both datasets and the best F1 on VinDr-Mammo, while remaining competitive on CMMD F1. We will also add an explicit note in §5.4.1 acknowledging the CMMD F1 result and framing the full framework as providing the best overall trade-off rather than uniformly dominating every metric. revision: yes

  2. Referee: Reference [26] is cited as 'MedSigLIP' throughout the text (e.g., §3, §5.1), but the bibliography entry is titled 'Medgemma technical report' (arXiv:2507.05201). MedGemma and MedSigLIP are distinct models. This mismatch must be resolved to ensure reproducibility, as the backbone's pretrained capacity fundamentally affects the interpretation of the baseline comparisons.

    Authors: The referee is correct, and we thank them for catching this error. Reference [26] currently points to the MedGemma technical report (arXiv:2507.05201), but the model actually used throughout all experiments is MedSigLIP, which is a distinct model with different pretrained weights. This is a bibliographic mistake: the correct reference for MedSigLIP should cite the appropriate MedSigLIP model card or publication. We will correct reference [26] to properly cite MedSigLIP and verify that all in-text references accurately reflect the model used. We will also add the specific model checkpoint identifier in §5.1 to ensure full reproducibility. No experiments need to be re-run, as the code and pretrained weights used are already those of MedSigLIP; only the citation was incorrect. revision: yes

  3. Referee: §5.1: The number and placement of fusion modules and prompt depth are tuned separately for each dataset using the validation set. With only two fusion blocks selected for VinDr-Mammo (e.g., layers 12 and 23 per §5.5), the risk of fitting to validation-set noise is non-trivial. The paper does not test cross-dataset transfer of these hyperparameters. The authors should discuss this overfitting risk and ideally report the sensitivity of the results to these specific architectural choices.

    Authors: This is a fair concern. The per-dataset tuning of fusion block placement and prompt depth does carry a risk of fitting to validation-set noise, particularly given the relatively small validation sets (e.g., 520 samples for VinDr-Mammo). We will address this in two ways in the revision. First, we will add a dedicated paragraph in §5.5 discussing the overfitting risk and noting that the sensitivity analysis in Figure 3 (varying prompt depth and number of fusion blocks) provides partial evidence that performance is not overly fragile to these choices — for instance, prompt depths 8, 12, and 16 yield comparable F1 on VinDr-Mammo 5-class, and both 1-block and 2-block fusion configurations are competitive on the binary task. Second, we will run a cross-dataset transfer experiment: the hyperparameters selected on VinDr-Mammo will be applied to CMMD and vice versa, and we will report the resulting performance. If the performance degradation is modest, this will provide evidence that the chosen configurations generalize reasonably; if not, we will report this honestly and discuss the implications. We acknowledge that a fully systematic sweep over all placement combinations would be ideal but is computationally expensive given the frozen backbone's forward cost; we will note this limitation explicitly. revision: partial

  4. Referee: §5.4.2 and Table 3: The comparison with existing methods is limited to DIVF [20], which is not a recent state-of-the-art method. The related work section discusses more recent and relevant multi-view transformer approaches (e.g., MaMVT [17], TransCorNet [24], BRAIxMVCCL [5]). The paper would be substantially strengthened by comparing against at least one of these recent baselines under identical backbone settings to properly contextualize the proposed framework's contributions.

    Authors: The referee is right that comparing only against DIVF [20] is insufficient to contextualize the contribution, especially given that the related work section discusses more recent and stronger multi-view methods. We will add at least one comparison against a recent multi-view transformer baseline. The most feasible candidate is BRAIxMVCCL [5], as it is the most directly comparable in task formulation (multi-view mammography classification with CC/MLO interaction). However, we note a practical difficulty: MaMVT [17], TransCorNet [24], and BRAIxMVCCL [5] were designed around different backbones (e.g., Swin Transformer, ResNet-based architectures) and are not natively built on a frozen MedSigLIP encoder. A fully fair comparison under 'identical backbone settings' would require reimplementing these methods on top of MedSigLIP, which is non-trivial. We will take the following approach: (a) reimplement at least one recent baseline (likely BRAIxMVCCL's cross-view interaction module) as an adapter on the same frozen MedSigLIP backbone, and (b) additionally report the published numbers from the original papers on VinDr-Mammo or CMMD if available, clearly noting the backbone difference. We will be transparent about which comparisons use the same backbone and which do not. If reimplementation proves infeasible within the revision timeline, we will at minimum add the published numbers with appropriate caveats and discuss the comparison qualitatively. revision: partial

Circularity Check

0 steps flagged

No circularity found — the framework's derivation chain is self-contained and does not reduce to its inputs by construction.

full rationale

The paper proposes a token-based dual-view fusion mechanism for mammogram classification and evaluates it on held-out test sets. Walking the derivation chain: (1) The fusion tokens are defined via bidirectional cross-attention (Eqs. for A_cc, A_mlo) followed by mean pooling and token insertion — this is an architectural design, not a definition that presupposes the evaluation metric. (2) The classification predictions are produced by a trained linear classifier on concatenated embeddings, not by any formula that inverts the fusion definition. (3) Hyperparameters (fusion block placement, prompt depth) are tuned on validation sets and results reported on separate test sets (Section 5.1, Table 1) — this is standard ML practice, not fitted-input-as-prediction circularity, since the test partition is disjoint from the validation partition. (4) No self-citations appear in the reference list; the authors (Pirsoltan, Babakordi, Mohammadi) do not cite their own prior work as load-bearing evidence. The cited methods (VPT [9], MaPLe [10], MedSigLIP [26]) are all external. (5) No uniqueness theorem is invoked. (6) The ablation studies (Tables 2, 4, 5, 6) compare the proposed mechanism against architecturally distinct alternatives (residual fusion, different pooling strategies, frozen vs. fine-tuned head), and the results differ — so the 'fusion tokens' contribution is not a renaming of a known result. The MedSigLIP/MedGemma citation ambiguity flagged by the reader is a correctness and reproducibility concern, not a circularity issue. The per-dataset hyperparameter tuning raises generalization risk but does not make the test-set results circular. Overall, the paper's central claims are supported by independent experimental evidence rather than by definitional reduction or self-citation chains.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The axiom ledger captures the key design choices and assumptions. The free parameters are mostly standard for prompt-learning and cross-attention methods, but the per-dataset tuning of fusion block placement is a notable source of potential overfitting. The invented entity (fusion tokens) is an architectural construct with falsifiable ablation evidence, not a postulated physical entity.

free parameters (5)
  • Deep prompt tokens = J tokens, one per layer for first J layers; J tuned per dataset (e.g., 12 for VinDr-5-class, 8 for VinDr-binary)
    Learnable parameters optimized during Stage 1; depth J is a hyperparameter tuned on validation set.
  • Cross-view fusion module weights = Multi-head cross-attention with 4 heads; trained in Stage 2
    Learnable parameters for Q/K/V projections in cross-attention blocks.
  • Fusion block placement = Dataset-specific (e.g., layers 12 and 23 for VinDr-binary; tuned on validation set)
    Hyperparameter selecting which transformer layers receive fusion modules; tuned per dataset.
  • Vision encoder head learning rate = 1e-5 (best from {frozen, 1e-7, 1e-5})
    Learning rate for fine-tuning the pooling head; selected via ablation.
  • Linear classifier weights = Not specified numerically
    Final classification layer trained on concatenated features.
axioms (4)
  • domain assumption MedSigLIP pretrained weights provide transferable representations for mammography
    The entire framework depends on the frozen backbone providing useful features; this is assumed, not proven, though it is a standard assumption in transfer learning.
  • ad hoc to paper Mean pooling of cross-attention outputs is the best aggregation strategy for fusion tokens
    Table 5 shows mean pooling outperforms max and attention pooling, but this is empirically determined on one dataset (VinDr-binary) and assumed to generalize.
  • domain assumption Patient-level dataset splitting prevents information leakage
    Section 4 states splits are patient-level; this is a standard and reasonable assumption.
  • domain assumption BI-RADS 2 = suspicious benign, BI-RADS 4+5 = suspicious malignancy
    Section 4 defines the binary grouping following [20]; this is a clinical assumption about BI-RADS categories.
invented entities (1)
  • Fusion tokens (t_cc, t_mlo) independent evidence
    purpose: Compact representations of cross-view information exchanged between CC and MLO views via cross-attention and mean pooling
    The fusion tokens are a structural component of the architecture, not a new physical entity. Their effectiveness is tested via ablation (Table 4: fusion tokens vs. residual fusion) and the results are falsifiable.

pith-pipeline@v1.1.0-glm · 16761 in / 2813 out tokens · 391752 ms · 2026-07-08T10:08:18.599618+00:00 · methodology

0 comments
read the original abstract

Accurate breast cancer classification from mammography requires effective integration of complementary information from craniocaudal (CC) and mediolateral oblique (MLO) views, which provide a more complete characterization of breast abnormalities. However, existing multi-view learning approaches typically rely on feature-level aggregation or single-stage cross-attention, which can entangle view-specific and shared representations and restrict interaction to limited network depths. To address these limitations, we propose a token-centric dual-view learning framework that unifies prompt-based adaptation and cross-view fusion within a frozen vision transformer backbone. The framework reformulates inter-view interaction as structured token-level communication, where dedicated fusion tokens explicitly encode bidirectional information exchange between CC and MLO views via cross-attention, serving as intermediate carriers of cross-view dependencies rather than relying on direct feature fusion. Unlike conventional methods that apply fusion at a single layer, fusion modules are inserted at multiple transformer depths, enabling progressive and repeated interaction across the encoder hierarchy. Fusion tokens are reintegrated into the token sequence and refined by subsequent transformer layers, facilitating hierarchical propagation of complementary information while preserving view-specific structure. Experiments on VinDr-Mammo and CMMD datasets demonstrate consistent improvements over linear probing, prompt-only adaptation, and conventional fusion baselines. On the VinDr-Mammo BI-RADS classification task, the framework achieves 50.40% F1-score and 0.8090 AUC, including a 0.10 AUC improvement over a dual-view fusion baseline in the binary setting. Ablation studies further validate the effectiveness of token-based fusion and multi-depth interaction design.

Figures

Figures reproduced from arXiv: 2607.06309 by Aysan Ghayouri Pirsoltan, Mohammad Reza Mohammadi, Shima Babakordi.

Figure 1
Figure 1. Figure 1: Overview of the proposed framework. Stage 1 learns view-consistent deep prompts, while Stage 2 introduces cross￾view fusion at selected transformer layers. Cross-View Fusion Block enables bidirectional information exchange between CC and MLO views. 𝑧 = Head(𝐻𝐾), 𝑧 ∈ ℝ𝐷. Each sample includes CC and MLO views, encoded through a single shared vision encoder with learnable prompt tokens: 𝑧𝑐𝑐 = 𝑓vision(𝐶𝐶), 𝑧𝑚𝑙… view at source ↗
Figure 2
Figure 2. Figure 2: Normalized confusion matrices on the VinDr-Mammo 5-class test set. Each row is normalized to sum to 100%. Compared with the linear probe baseline using average fusion (MAE = 0.5575), the proposed method achieves a substantially lower mean absolute error (MAE = 0.3820) and exhibits stronger diagonal dominance, indicating improved classification accuracy and fewer ordinal prediction errors across BI-RADS cat… view at source ↗
Figure 3
Figure 3. Figure 3: (a) Effect of prompt depth on F1-score for the VinDr-Mammo 5-class classification task. (b) AUC-ROC performance on the VinDr-Mammo binary classification task for different numbers of cross-view fusion blocks, where 𝑁 denotes the number of fusion blocks. Summary. The ablation results show that each component contributes to the overall performance. The fusion￾token mechanism consistently improves results ove… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative interpretability results obtained using occlusion-based sensitivity analysis [29]. From left to right, the columns present the original craniocaudal (CC) image, the CC overlay, the CC sensitivity map, the original mediolateral oblique (MLO) image, the MLO overlay, and the MLO sensitivity map. Regions with warmer colors correspond to image areas whose occlusion leads to a larger reduction in the… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 5 internal anchors

  1. [1]

    Mammo-bench:Alarge-scalebenchmarkdatasetofmammographyimages,in:InternationalConference on Computational Advances in Bio and Medical Sciences, Springer

    Bhole,G.,Suba,S.,Parekh,N.,2025. Mammo-bench:Alarge-scalebenchmarkdatasetofmammographyimages,in:InternationalConference on Computational Advances in Bio and Medical Sciences, Springer. pp. 144–156

    work page 2025

  2. [2]

    An online mammography database with biopsy confirmed types

    Cai, H., Wang, J., Dan, T., Li, J., Fan, Z., Yi, W., Cui, C., Jiang, X., Li, L., 2023. An online mammography database with biopsy confirmed types. Scientific Data 10, 123

    work page 2023

  3. [3]

    Convolutional networks and transformers for mammography classification: an experimental study

    Cantone, M., Marrocco, C., Tortorella, F., Bria, A., 2023. Convolutional networks and transformers for mammography classification: an experimental study. Sensors 23, 1229

    work page 2023

  4. [4]

    Adaptformer: Adapting vision transformers for scalable visual recognition

    Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P., 2022a. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664–16678

    work page

  5. [5]

    Chen, Y., Wang, H., Wang, C., Tian, Y., Liu, F., Liu, Y., Elliott, M., McCarthy, D.J., Frazer, H., Carneiro, G., 2022b. Multi-view local co-occurrence and global consistency learning improve mammogram classification generalisation, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 3–13

    work page

  6. [6]

    Interpretable breast cancer identification by multi-view synergistic feature fusion interaction in dense breast tissue

    Dai, G., Wang, C., Dai, D., Tang, Q., Zhang, Y., Chen, H., 2026. Interpretable breast cancer identification by multi-view synergistic feature fusion interaction in dense breast tissue. Information Fusion 125, 103446

    work page 2026

  7. [7]

    Efficient breast cancer mammograms diagnosis using three deep neural networks and term variance

    Elkorany, A.S., Elsharkawy, Z.F., 2023. Efficient breast cancer mammograms diagnosis using three deep neural networks and term variance. Scientific Reports 13, 2663

    work page 2023

  8. [8]

    Lora: Low-rank adaptation of large language models

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al., 2022. Lora: Low-rank adaptation of large language models. Iclr 1, 3

    work page 2022

  9. [9]

    Visual prompt tuning, in: European conference on computer vision, Springer

    Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N., 2022. Visual prompt tuning, in: European conference on computer vision, Springer. pp. 709–727

    work page 2022

  10. [10]

    Maple: Multi-modal prompt learning, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S., 2023. Maple: Multi-modal prompt learning, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 19113–19122

    work page 2023

  11. [11]

    Adam: A Method for Stochastic Optimization

    Kingma, D.P., Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  12. [12]

    Stethonet:robustbreastcancermammography classification framework

    Lamprou,C.,Katsikari,K.,Rahmani,N.,Hadjileontiadis,L.J.,Seghier,M.,Alshehhi,A.,2024. Stethonet:robustbreastcancermammography classification framework. IEEE Access . Preprint submitted to Elsevier on July 2, 2026Page 12 of 13

    work page 2024

  13. [13]

    Multi-view cross attention for guiding surgical planning in rotator cuff repairs with limited mri data

    Lee, Y.O., Lee, H., Kim, J.H., 2025. Multi-view cross attention for guiding surgical planning in rotator cuff repairs with limited mri data. IEEE Access 13, 215771–215785

    work page 2025

  14. [14]

    An interpretable cnn-based model for mass classification in mammography

    Li, G., Zhou, M., Fu, Y., Alam, N., Denton, E., Zwiggelaar, R., 2025. An interpretable cnn-based model for mass classification in mammography. Knowledge-Based Systems 316, 113372

    work page 2025

  15. [15]

    Actlikearadiologist:towardsreliablemulti-viewcorrespondencereasoning for mammogram mass detection

    Liu,Y.,Zhang,F.,Chen,C.,Wang,S.,Wang,Y.,Yu,Y.,2021a. Actlikearadiologist:towardsreliablemulti-viewcorrespondencereasoning for mammogram mass detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 5947–5961

    work page

  16. [16]

    Swin transformer: Hierarchical vision transformer using shifted windows, in: Proceedings of the IEEE/CVF international conference on computer vision, pp

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B., 2021b. Swin transformer: Hierarchical vision transformer using shifted windows, in: Proceedings of the IEEE/CVF international conference on computer vision, pp. 10012–10022

    work page

  17. [17]

    Mammographyclassificationwithmulti-view deep learning techniques: Investigating graph and transformer-based architectures

    Manigrasso,F.,Milazzo,R.,Russo,A.S.,Lamberti,F.,Strand,F.,Pagnani,A.,Morra,L.,2025. Mammographyclassificationwithmulti-view deep learning techniques: Investigating graph and transformer-based architectures. Medical Image Analysis 99, 103320

    work page 2025

  18. [18]

    TransReg: Cross-transformer as auto-registration module for multi-view mammogram mass detection

    Nguyen, H.C., Phan, C., Pham, H.H., 2023a. Transreg: Cross-transformer as auto-registration module for multi-view mammogram mass detection. arXiv preprint arXiv:2311.05192

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Vindr-mammo: A large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography

    Nguyen, H.T., Nguyen, H.Q., Pham, H.H., Lam, K., Le, L.T., Dao, M., Vu, V., 2023b. Vindr-mammo: A large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography. Scientific Data 10, 277

    work page

  20. [20]

    Nguyen, T.H., Kha, Q.H., Truong, T.N.T., Lam, B.T., Ngo, B.H., Dinh, Q.V., Le, N.Q.K., 2023c. Towards robust natural-looking mammography lesion synthesis on ipsilateral dual-views breast cancer analysis, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pp. 2564–2573

    work page

  21. [21]

    Mv-trams: An efficient tumor region-adapted mammography synthesisundermulti-viewdiagnosis

    Nguyen, T.H., Lam, B.T., Truong, T.N.T., Duong, D.T., Dinh, Q.V., 2026. Mv-trams: An efficient tumor region-adapted mammography synthesisundermulti-viewdiagnosis. BiomedicalSignalProcessingandControl123,110509. URL:https://www.sciencedirect.com/ science/article/pii/S1746809426010633, doi:https://doi.org/10.1016/j.bspc.2026.110509

    work page doi:10.1016/j.bspc.2026.110509 2026

  22. [22]

    Weakly supervised learning for breast cancer prediction on mammograms in realistic settings

    Pathak, S., Schlötterer, J., Geerdink, J., Vijlbrief, O.D., van Keulen, M., Seifert, C., 2023. Weakly supervised learning for breast cancer prediction on mammograms in realistic settings

    work page 2023

  23. [23]

    Breast cancer diagnosis in two-view mammography using end-to-end trained efficientnet-based convolutional network

    Petrini, D.G., Shimizu, C., Roela, R.A., Valente, G.V., Folgueira, M.A.A.K., Kim, H.Y., 2022. Breast cancer diagnosis in two-view mammography using end-to-end trained efficientnet-based convolutional network. Ieee access 10, 77723–77731

    work page 2022

  24. [24]

    Transformer-guidedcross-viewcorrelationnetworkformammogramclassification,in: 2025 IEEE 9th International Conference on Information and Communication Technology (CICT), IEEE

    Ramakrishnan,K.,Mohideen,K.,Reddy,A.A.,2025. Transformer-guidedcross-viewcorrelationnetworkformammogramclassification,in: 2025 IEEE 9th International Conference on Information and Communication Technology (CICT), IEEE. pp. 1–6

    work page 2025

  25. [25]

    Deep learning in mammography images segmentation and classification: Automated cnn approach

    Salama, W.M., Aly, M.H., 2021. Deep learning in mammography images segmentation and classification: Automated cnn approach. Alexandria Engineering Journal 60, 4701–4709

    work page 2021

  26. [26]

    MedGemma Technical Report

    Sellergren, A., Kazemzadeh, S., Jaroensri, T., Kiraly, A., Traverse, M., Kohlberger, T., Xu, S., Jamil, F., Hughes, C., Lau, C., et al., 2025. Medgemma technical report. arXiv preprint arXiv:2507.05201

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    Spak,D.A.,Plaxco,J.,Santiago,L.,Dryden,M.,Dogan,B.,2017.Bi-rads®fifthedition:Asummaryofchanges.Diagnosticandinterventional imaging 98, 179–190

    work page 2017

  28. [28]

    Van Tulder, G., Tong, Y., Marchiori, E., 2021. Multi-view analysis of unregistered medical images using cross-view transformers, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 104–113

    work page 2021

  29. [29]

    Visualizingandunderstandingconvolutionalnetworks,in:Europeanconferenceoncomputervision,Springer

    Zeiler,M.D.,Fergus,R.,2014. Visualizingandunderstandingconvolutionalnetworks,in:Europeanconferenceoncomputervision,Springer. pp. 818–833

    work page 2014

  30. [30]

    Tip-Adapter: Training-free CLIP-Adapter for Better Vision-Language Modeling

    Zhang,R.,Fang,R.,Zhang,W.,Gao,P.,Li,K.,Dai,J.,Qiao,Y.,Li,H.,2021. Tip-adapter:Training-freeclip-adapterforbettervision-language modeling. arXiv preprint arXiv:2111.03930

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  31. [31]

    BiomedCLIP: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs

    Zhang,S.,Xu,Y.,Usuyama,N.,Xu,H.,Bagga,J.,Tinn,R.,Preston,S.,Rao,R.,Wei,M.,Valluri,N.,etal.,2023. Biomedclip:amultimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  32. [32]

    Learning to prompt for vision-language models

    Zhou, K., Yang, J., Loy, C.C., Liu, Z., 2022. Learning to prompt for vision-language models. International journal of computer vision 130, 2337–2348. Preprint submitted to Elsevier on July 2, 2026Page 13 of 13

    work page 2022