List of Papers Browse by Subject Areas Author List
Abstract
Benign–malignant classification in automated breast ultrasound (ABUS) is challenging due to heterogeneous voxel spacing, anisotropic artifacts, and weakly localized malignancy cues distributed across 3D volumes. We propose MSQG-3DNet, a standardized-lattice CNN–Transformer hybrid for lesion-level ROI-based 3D ABUS classification, coupling multi-scale volumetric representation learning with query-guided global evidence aggregation. To reduce inter-case intensity and input-size variability, lesion-centered ROI volumes are intensity-normalized and standardized to a fixed 48 × 48 × 48 lattice. The network employs an SE-enhanced dilated 3D residual encoder to capture multi-scale morphology, followed by dual-scale tokenization with hybrid 3D positional encoding, including coordinate-based encoding, sinusoidal encoding, and scale embedding, to preserve spatial and scale awareness under lattice standardization. LayerScale-stabilized Transformer blocks model long-range dependencies, and a lightweight cross-attention Query-Decoder extracts task-relevant evidence. Finally, a gated fusion integrates complementary global descriptors for malignancy prediction. On an in-house ABUS dataset (8,627 training and 859 test volumes), MSQG-3DNet achieves a test AUC of 96.23%. External evaluation on the public TDSC-ABUS benchmark yields the highest threshold-independent AUC of 71.55% among compared methods, indicating improved benign–malignant discrimination under domain shift, while highlighting the need for operating threshold calibration in clinical use. Code available at: https://github.com/Ferry-Rain-Way/MSQG-3DNet.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2113_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Ferry-Rain-Way/MSQG-3DNet
Link to the Dataset(s)
https://tdsc-abus2023.grand-challenge.org/
BibTex
@InProceedings{LiuZef_MSQG3DNet_MICCAI2026,
author = { Liu, Zefeng AND Duan, Yaofei AND Yang, Zhikai AND Xu, Xiaojing AND Gao, Jie AND Huang, Jiaju AND Sun, Yue AND Yang, Kaiwen AND Bao, Lingyun AND Tan, Tao},
title = { { MSQG-3DNet: Multi-scale Token Selection with Query-Decoder and Gated Fusion for 3D Breast Tumor Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes MSQG-3DNet, a 3D ABUS benign-versus-malignant classifier that combines a residual 3D encoder, multi-scale tokenization with hybrid 3D positional encoding, informative token selection, a query-guided decoder, and gated fusion of global descriptors. The submission also reports both a large internal hold-out evaluation and an external evaluation on TDSC-ABUS.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper addresses a clinically relevant 3D ABUS classification problem where weak malignancy cues can be spatially diffuse and difficult to capture with purely local models. The proposed architecture is reasonably coherent: token pruning, the query decoder, and gated fusion are not presented as isolated tricks, but as a linked pipeline for reducing redundancy and aggregating sparse global evidence. The internal study is fairly substantial for this application area, with 9,486 annotated volumes, a dedicated 859-case test split, progressive ablations, and comparison against prior ABUS-oriented baselines. I also view the external TDSC-ABUS experiment as a positive aspect, because the authors at least attempt to test cross-dataset behavior rather than reporting only single-center results.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The main concern is generalization. The external TDSC-ABUS test set contains only 49 cases, and on that set the paper reports the best AUC but only 44.83% sensitivity, so the current evidence is not strong enough to support a broad robustness claim under domain shift. A second issue is the aggressive preprocessing: every volume is normalized and standardized to 48x48x48, yet the paper is motivated by subtle and weakly localized malignancy cues, so the effect of this resampling choice on fine morphology should be analyzed rather than treated as neutral. The comparison set is also limited to two re-implemented competing methods, which makes it difficult to judge how much of the gain comes from the proposed architectural contributions versus the authors’ overall training recipe. Finally, the operating-point story needs more discussion. On the internal set, the final model attains 100% specificity but lower sensitivity than one ablation; on the external set, it again improves specificity at the cost of markedly lower sensitivity. That trade-off is clinically meaningful and should be discussed explicitly, not only through AUC.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
For rebuttal, the most useful additions would be: (1) clarify how thresholds were selected and whether the method can recover a higher-sensitivity operating point without collapsing specificity; (2) quantify how sensitive the results are to the fixed 48x48x48 resampling choice; and (3) explain the external performance drop more carefully, including whether it is mainly due to acquisition differences, annotation differences, or class-composition mismatch. I would also recommend softening the domain-robustness language unless stronger multi-center evidence is available.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I place this paper slightly above the acceptance threshold. The method is technically coherent, the internal evaluation is stronger than average for this application area, and the external test - although limited - is still more informative than having no cross-dataset evaluation at all. My enthusiasm is moderated by the small external cohort, the large internal-to-external performance gap, and the limited breadth of comparisons. Overall, I see this as a credible borderline accept rather than a clearly strong paper.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the authors’ rebuttal, I appreciate the detailed clarifications and the generally constructive responses to the reviewers’ concerns. The authors adequately clarified the intended clinical scope of the method as a lesion-level characterization framework rather than a complete CAD system, which makes the experimental setting more reasonable. The rebuttal also improved the explanation of the multi-scale token interaction mechanism and clarified that the proposed framework is not merely based on naive token concatenation.
I also appreciate the additional discussion regarding threshold selection and the SEN/SPE trade-off. The authors explained the use of Youden’s index and acknowledged the clinical importance of balancing sensitivity and specificity. In addition, the discussion on the external TDSC-ABUS evaluation and the challenges caused by missing voxel spacing information helps contextualize the observed performance drop under domain shift.
That said, some limitations remain only partially resolved. In particular, the external validation cohort is still relatively small, and the robustness claims should remain conservative. The concern regarding the potential impact of aggressive 48×48×48 resampling on subtle morphological cues was not fully addressed through additional analysis. The comparison breadth also remains somewhat limited. However, I acknowledge that these limitations are partly related to the scarcity of publicly available and standardized 3D ABUS benchmarks.
Overall, despite these remaining weaknesses, I still find the work technically coherent and experimentally stronger. The internal evaluation is substantial, the ablation study is reasonably comprehensive, and the authors made a meaningful effort to evaluate cross-dataset behavior. Therefore, I maintain my borderline positive acceptance.
Review #2
- Please describe the contribution of the paper
The paper proposes a CNN–Transformer hybrid framework for 3D ABUS benign–malignant classification, combining multi-scale feature extraction with token-based global modeling. It introduces a token selection mechanism, a query-based aggregation module, and a gated fusion strategy to integrate multiple global representations for final prediction.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Addresses an important problem in 3D breast ultrasound (ABUS) classification. Proposes a coherent CNN–Transformer hybrid framework combining multi-scale feature extraction with global modeling. Introduces a structured pipeline (token selection, query-based aggregation, and fusion) for feature integration. Provides ablation studies to analyze the contribution of different components. The paper is clearly written and well-organized, making the methodology easy to follow.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The task is formulated as lesion-centered classification, where the input volumes are already cropped around annotated lesions. While this setting is valid for diagnostic support, it significantly simplifies the problem by removing the need for lesion detection. As a result, the practical applicability of the proposed method in a fully automated clinical pipeline (e.g., screening) may be limited. 2.The multi-scale tokens are simply concatenated and processed with full self-attention, which is a rather brute-force design. This introduces unnecessary interactions across all tokens and is not an efficient or well-structured way to model cross-scale relationships. 3.The model consistently exhibits low sensitivity and high specificity across both datasets, indicating a strong bias toward predicting the benign class. This behavior can be attributed to two factors. First, the in-house dataset is highly imbalanced, with significantly more benign than malignant cases, which may encourage the model to favor the majority class. Notably, this issue becomes even more pronounced on the smaller external dataset, which is relatively balanced, where the sensitivity drops substantially, suggesting poor generalization to malignant cases. Second, the model primarily relies on global representation learning, while malignancy classification often depends on subtle and localized features. This suggests that the proposed architecture may be insufficient in capturing fine-grained local cues critical for detecting malignant lesions.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I lean toward a weak reject. While the paper presents a well-structured CNN–Transformer hybrid framework and provides clear experimental analysis with ablation studies, the overall novelty is limited, as the method mainly combines existing components rather than introducing a fundamentally new design. In addition, the experimental evaluation is not sufficiently comprehensive, with only a small number of baselines and limited comparison to recent relevant methods. More importantly, the model consistently shows low sensitivity and high specificity, indicating a bias toward benign predictions, which is problematic for clinical applications. This issue appears to be related to both dataset imbalance and the model’s reliance on global representations, which may be insufficient for capturing subtle local malignant cues. Overall, while the work is technically sound, stronger justification of its design choices and more convincing empirical validation are needed.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes MSQG-3DNet, a CNN-Transformer hybrid architecture for benign-malignant tumor classification in 3D Automated Breast Ultrasound (ABUS). The core methodology addresses the challenges of heterogeneous voxel spacing and weak, dispersed malignancy cues. The authors introduce a pipeline consisting of a residual 3D CNN with Multi-Scale Attention Fusion (MSAF) for local feature extraction, followed by a Transformer encoder. To handle the computational complexity of 3D tokens, the authors incorporate an Informative Token Pruning module, a Query-Decoder for task-focused global evidence aggregation, and a Gated Fusion mechanism to adaptively combine multiple global descriptors (class token, CNN features, pooled tokens, and query tokens). The model is evaluated on a notably large in-house dataset (9,486 volumes) and tested for domain shift on the public TDSC-ABUS dataset.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Extensive In-house Dataset: The use of an in-house dataset comprising 9,486 3D ABUS volumes is highly commendable. Annotating and utilizing 3D medical datasets of this scale is rare and provides a strong foundation for the in-house evaluation.
Logical Architectural Design: The integration of CNNs and Transformers is well-motivated. The specific combination of Multi-Scale Attention Fusion, Token Pruning, and Query-Decoder effectively addresses the balance between local textural feature extraction and computational efficiency for 3D volumes.
Thorough Ablation Study: Table 2 and Figure 2 provide a clear, progressive validation of the proposed components. The training dynamics demonstrate that the token pruning and query-guided aggregation successfully stabilize the optimization process.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Unacceptable Clinical Sensitivity on External Data: While the model achieves the highest AUC on the external TDSC-ABUS dataset, its Sensitivity drops drastically to 44.83% (Table 4). In the context of breast cancer screening and diagnosis, missing more than half of the malignant tumors (false negatives) is a critical failure. The model becomes overly conservative (heavily favoring the benign class) under domain shift, rendering the “improved robustness” claim highly questionable from a clinical perspective.
Unrealistic Inference Assumptions (Cropping): The paper states that “lesion-centered 3D bounding-box crops were generated programmatically” based on radiologist annotations. It is entirely unclear how the model operates during real-world inference. Does it require a radiologist or a separate detection network to provide a perfect 3D bounding box first? If so, this limits the method’s fully automated CAD utility and makes the pipeline incomplete.
Limited Methodological Novelty: The individual components—multi-scale CNN backbones, token pruning, query-based cross-attention, and gated fusion—are directly adapted from existing standard computer vision literature without significant domain-specific algorithmic innovation.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend a “Weak Reject.” While the scale of the in-house dataset is highly impressive and the network architecture is logically constructed, several fundamental flaws prevent acceptance in its current form. First and foremost, the severe drop in Sensitivity (44.83%) on the external dataset indicates the model is highly biased toward predicting benign cases under domain shift, making the proposed algorithm clinically unviable in its current state. Second, the assumption of having pre-existing lesion-centered 3D crops for classification is unrealistic for an end-to-end automated system, and the paper fails to explain how this is handled during inference. Finally, the aggressive downsampling to a 48x48x48 grid risks destroying the exact morphological cues the network is supposed to detect. Unless the authors can provide compelling clarifications regarding the inference pipeline and a solution (or at least a rigorous analysis) for the catastrophic drop in external sensitivity in their rebuttal, the paper falls short of the MICCAI acceptance threshold.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #4
- Please describe the contribution of the paper
This paper proposes MSQG-3DNet, a CNN–Transformer hybrid framework for benign–malignant classification in 3D automated breast ultrasound (ABUS). The main contribution lies in a progressive multi-stage architecture that integrates local volumetric feature learning with global token-based reasoning. Specifically, the method introduces: (1) a multi-scale tokenization with hybrid 3D positional encoding to preserve spatial consistency after resampling, (2) an informative token selection mechanism to reduce redundancy and focus on lesion-relevant regions, (3) a query-guided decoder for task-oriented global evidence aggregation via cross-attention, and (4) an adaptive gated fusion module to combine multiple global descriptors for robust prediction.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Well-designed CNN–Transformer hybrid architecture The method effectively combines 3D CNNs for local feature extraction with Transformer-based global reasoning, which is well suited for handling dispersed and subtle malignancy cues in volumetric ABUS data. Coherent and task-driven module integration The proposed Token Selection → Query-Decoder → Gated Fusion pipeline forms a logical and effective framework for reducing redundancy, aggregating global evidence, and improving prediction robustness. Task-specific adaptation for 3D data The hybrid 3D positional encoding is well motivated and helps preserve spatial consistency under resampling, addressing key challenges in volumetric medical imaging. Relatively comprehensive experimental validation The paper includes ablation studies and external dataset evaluation, providing useful insights into component contributions and demonstrating a degree of generalization under domain shift.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Limited methodological novelty The framework mainly integrates existing techniques (e.g., CNN–Transformer hybrids , token pruning , and query-based cross-attention ), with limited fundamentally new components, although the overall design is coherent. 2.Generalization could be improved Performance drops on the external dataset (AUC 71.55), and the relatively low sensitivity suggests remaining challenges under domain shift. 3.Limited comparison with recent baselines The evaluation lacks comparison with more recent or stronger Transformer-based methods, which would better position the contribution. 4.Missing efficiency analysis The paper does not report model complexity (e.g., parameters, inference time), making it difficult to assess the performance–efficiency trade-off.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
none
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
none
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a well-structured CNN–Transformer hybrid framework that is thoughtfully adapted to the challenges of 3D ABUS data.However, the methodological novelty is somewhat limited, as many components build upon existing ideas such as CNN–Transformer hybrids, token pruning, and query-based cross-attention. In addition, the performance drop on the external dataset and the relatively low sensitivity indicate that generalization under domain shift remains an open issue. The evaluation could also be strengthened by including more recent and competitive baselines, as well as analysis of computational efficiency. Overall, the paper demonstrates solid engineering and practical value, but would benefit from stronger novelty and more comprehensive validation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal satisfactorily clarifies several reviewer concerns, particularly regarding the SEN–SPE trade-off and the intended lesion-level classification setting. While concerns about external generalization and limited methodological novelty remain, I believe the work is technically sound and the rebuttal improves the overall clarity and positioning of the paper.
Author Feedback
We thank all reviewers (R) for their feedback. We have clarified the concerns and will improve the writing accordingly.
1.SEN, SPE, and clinical use (R1,2,3,4) Thank you for the valuable comment. SEN and SPE were first computed using the default threshold of 0.5.To better balance them, we further used Youden’s index to select the optimal threshold. This gives 68.97% SEN and 60.00% SPE on this dataset. MSQG-3DNet still achieves the highest threshold-independent AUC on both internal and external datasets, showing better benign–malignant discrimination. We will revise the manuscript to discuss this SEN–SPE trade-off.
2.External validation (R1,2,3,4) Public, well-annotated 3D ABUS datasets are still very limited. Our external set comes from the MICCAI TDSC-ABUS Challenge, which is, to our knowledge, the only public benchmark for 3D ABUS analysis, and was also used in SAMASK-CLTR (MICCAI 2025). A key difficulty is that TDSC-ABUS2023 does not provide voxel spacing information, making precise physical-scale normalization difficult and increasing domain shift. All methods showed an AUC drop on this set, suggesting common domain and small-sample challenges. Still, MSQG-3DNet achieved the best external AUC, outperforming others by 7.24% and 7.76%, supporting its relative robustness under the available external setting.
3.Clarifying ROI-Based Classification Scope (R1,2,3) We did not claim MSQG-3DNet as a full CAD system. It focuses on lesion-level benign–malignant classification after lesion localization. The ROI can be provided by radiologists or an upstream detection/segmentation module, so MSQG-3DNet serves as a downstream lesion characterization model. This setting is clinically meaningful and widely used in breast imaging, as shown by HoVer-Trans (IEEE TMI 2023), FedBCD (IEEE TMI 2025), and GMIC (MedIA 2021).
4.Limited innovation (R2,3,4) We respectfully disagree. MSQG-3DNet is not simple module stacking. Each module has a clear role and gain. The backbone raises AUC from 86.35% to 95.10% and SEN from 51.01% to 78.39%. MSAF raises SEN to 81.56%. With TFM, MSQG-3DNet reaches the best AUC 96.23%, ACC 91.27%, and SPE 100.00%. The novelty lies in a task-specific 3D ABUS design: multi-scale tumor feature extraction, spatial-scale-aware token modeling, redundant token pruning, query-guided evidence aggregation, and adaptive descriptor fusion. Overall, it improves AUC by 9.88%, ACC by 11.06%, and SEN by 27.38% over V1. 5.Insufficient baseline comparison (R1,4) Publicly reproducible 3D ABUS classification baselines remain limited because well-annotated public ABUS datasets and standardized benchmarks are scarce. We therefore compared MSQG-3DNet with two representative methods, SAMASK-CLTR(MICCAI 2025) and 3DGAUnet (Cancers 2023), under the same data split and optimization settings to ensure a fair evaluation. The results show that MSQG-3DNet achieves stronger overall performance.
6.Coarse multi-scale token fusion (R2) We clarify that MSQG-3DNet does not perform simple token concatenation alone. Although multi-scale tokens are initially merged into one sequence, each token is augmented with hybrid 3D positional encoding, including spatial coordinates, sinusoidal encoding, and scale embedding, enabling spatial- and scale-aware Transformer interaction. Informative token pruning then suppresses redundant background tokens, while the Query-Decoder and adaptive gated fusion further aggregate lesion-relevant multi-scale evidence.
7.Efficiency analysis (R4) We agree and will add an efficiency analysis. MSQG-3DNet is designed for accurate 3D ABUS classification rather than real-time inference, with lesion-centered ROIs and token pruning reducing redundant computation. It has 165.91M parameters and requires 7.55 GMACs / 15.10 GFLOPs, achieving better performance under comparable computational cost.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Review comments are mixed. However, all the reviewers have expressed their concerns on the performance drop on the external set, along with many others. The authors are invited to addres the reviewer concerns as many as possible and clarify the paper’s arguments in this rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Two post-rebuttal comments agreed on acceptance of the paper. The rebuttal has clearly explained the external validation results and discussed balancing the sensitivity and specificity after applying the Youden Index. The crop-then-classify approach, while questioned by some reviewers, is well-established in the lesion-classification literature, and the authors’ use of it is consistent with some of the prior practice. The AC acknowledges the novelty of the classification method itself. In summary, the AC is convinced that the authors have addressed most of the main weaknesses and hence lean towards acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Reviewers agreed on the rebuttal and merits of the paper, the breast tumor segmentation topic is significant and rarely explored, and the paper provides insights on the task.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have addressed several concerns, the model performance on external validation need further discussion.