List of Papers Browse by Subject Areas Author List
Abstract
Neural Architecture Search (NAS) has shown significant potential in designing deep neural networks for medical image segmentation. However, even emerging training-free NAS frameworks often incur substantial computational costs and lengthy search times. To address the critical challenges of computational efficiency and architecture interpretability, the paper proposes a compact training-free NAS framework based on an Alternating Evolution Game (AEG-cTFNAS). The proposed method alternates the search and contribution evaluation of the encoder and decoder within the UNet architecture via alternating games. It employs a truncated normal distribution for compact encoding, sampling, and updating to minimize computational overhead, while Bayesian inference is utilized to estimate the contribution of each block, adaptively adjusting the search strategy and facilitating process visualization. Experimental results on two benchmark datasets reveal that AEG-cTFNAS outperforms both manually designed architectures and NAS-based algorithms, underscoring its efficacy and potential on medical image segmentation. Code is available at https://github.com/spcity/AEG-cTFNAS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1154_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/spcity/AEG-cTFNAS
Link to the Dataset(s)
ACDC dataset: https://www.creatis.insa-lyon.fr/Challenge/acdc/
EndoVis2018 dataset: https://opencas.dkfz.de/endovis/challenges/2018/
BibTex
@InProceedings{SunXia_Compact_MICCAI2025,
author = { Sun, Xiaoxue and Wang, Hongpeng and Song, Pei-Cheng},
title = { { Compact Training-free NAS with Alternating Evolution Game for Medical Image Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {107 -- 117}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces AEG-cTFNAS, a TFNAS framework for medical image segmentation. It combines truncated normal distribution-based sampling, an alternating evolution game to coordinate encoder-decoder updates, and Bayesian inference to estimate module contributions without extra cost. The method reduces search time and achieves good performance on benchmark datasets.
- 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.
- Contributions are clear and each contribution is valid in the ablation
- A novel encoder-decoder interaction game contribution evaluation strategy
- The search efficiency of the whole TFNAS is improved by Bayesian inference and truncated normal distribution, and the technique is soundly
- 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 proposed method focuses more on the efficiency improvement of TFNAS by designing efficient and interpretable contribution evaluation strategies and is less or largely irrelevant to MICCAI or the medical community. 2.The structure of the paper is not clear, especially last 2 paragraphs in introduction section, which further makes it difficult to identify the medical contribution of the paper.
- Confusion about dataset selection. I think ACDC and Endovis seems not to be a common benchmark to compare network structure, comparing with BTCV etc. Is it because the design of AEG-cTFNAS is relevant to both medical data scenarios?
- Comparison methods. To a certain extent, the contribution of the work lies in choosing a most efficient network structure, so I recommend please try to compare it with nnUnet.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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/2025/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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I think the main drawback of the article is that the medical implications of this article are very orthogonal to the technical design compared to existing TFNAS methods in the medical community such as [20] MedNAS. At the same time, I am confused about the choice of comprasion dataset.
- 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 author addressed my concerns
Review #2
- Please describe the contribution of the paper
The paper introduces AEG-cTFNAS, a novel compact training-free neural architecture search framework for medical image segmentation. Its main contributions include the use of a truncated normal distribution for compact encoding and efficient sampling, which reduces computational overhead. It proposes an alternating evolution game mechanism to dynamically alternate the search and update of encoder and decoder modules based on their contributions, thereby enhancing search efficiency. Additionally, it leverages Bayesian inference to estimate module contributions without additional evaluations, enabling real-time visualization and strategy adjustment. These innovations collectively reduce search costs and improve performance, achieving state-of-the-art results on ACDC datasets.
- 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 architecture proposed in this paper is innovative, setting it apart from most existing NAS approaches. Instead of conventional encoding methods, this paper employs a truncated normal distribution to achieve efficient encoding. Its use of Bayesian inference (BI) is also novel. By constructing a dataset and applying a logistic regression model, BI can instantly assess each module’s impact on overall performance. These statistical techniques are pioneering and captivating within the NAS field.
- 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 experimental section compares too few NAS methods, and NASUNet is somewhat outdated. More NAS methods should be contrasted to demonstrate the method’s effectiveness.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/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.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The neural architecture search method proposed in this paper is highly innovative and significantly different from previous related work. Instead of conventional encoding methods, this paper employs a truncated normal distribution for efficient encoding. The use of Bayesian inference, which constructs datasets and applies logistic regression models to assess each module’s impact on overall performance in real-time, is also novel. These statistical methods offer many new ideas for the development of the NAS field.
- 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 authors have addressed some of my concerns. So I recommend its acceptance
Review #3
- Please describe the contribution of the paper
The paper introduces AEG-cTFNAS, a compact, training-free NAS framework based on an Alternating Evolution Game strategy for medical image segmentation. It searches encoder and decoder structures separately and integrates truncated normal distribution sampling with Bayesian inference to guide architecture evaluation and enable visualization of the search process.
- 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.
(1)The authors provide open-source code, which supports transparency and reproducibility. (2)The method is evaluated on two benchmark datasets, with multiple baselines and settings. (3)The inclusion of visualizations for the NAS search in fig3 enhances interpretability, which is a valuable aspect in the medical imaging domain.
- 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.While the alternating evolution game is a novel formulation, the paper does not sufficiently position itself relative to existing training-free NAS approaches. A detailed discussion comparing AEG-cTFNAS with these works or those reported in the experiments would clarify its contribution and significance. 2.The paper only compares with NASUNet (2021) and MedNAS (2024). Broader and more up-to-date comparisons with strong NAS baselines—both training-free and differentiable—are necessary to establish the effectiveness of the proposed approach.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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/2025/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.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper presents a well-structured, interpretable NAS framework for medical image segmentation. The alternating encoder-decoder evolution, use of Bayesian inference, and compact sampling strategy are interesting design choices that contribute to the method’s practicality. The visualization of the search process is a valuable addition, and the experimental results are reasonably thorough.
However, the paper falls short in clearly contextualizing its novelty relative to recent training-free NAS literature and lacks comprehensive SOTA comparisons. Despite these limitations, the overall contribution is meaningful, especially for efficient and interpretable NAS in the medical domain. Therefore, I lean towards a weak accept.
- 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
Author Feedback
We are grateful for the insightful comments from the reviewers. Your expertise and attention to detail have significantly enhanced our work. We appreciate the reviewers’ recognition of the innovation and contributions of our architecture and methods [R1,R2,R3], the medical value of the visualization capabilities [R2], and algorithm’s role in improving TFNAS efficiency [R1,R3]. This paper presents a TFNAS-based medical image segmentation method offering efficient search (via compact encoding) and enhanced safety (from interpretable visualizations), aimed at future integration into surgical robots for surgical planning and navigation. We now categorically address the main points raised.
- Comparison with more NAS, TFNAS methods, and nnU-Net [R1,R2,R3]: While broader SOTA comparisons are valuable, the paper focuses on AEG-CTFNAS’s core innovations within the TFNAS framework: a faster search algorithm and distinctive visualization capabilities. Table 1 already validates its advantages within this defined scope. Preliminary results also indicate AEG-CTFNAS’s competitiveness against established methods like nnU-Net, Auto-DeepLab, and Bix-Net, with detailed findings to be presented in the future paper. More comprehensive comparative studies are deferred to future research.
- Contribution and significance compared to other TFNAS methods [R2]: This paper primarily focuses on the efficiency of the search algorithm. We observe that in the TFNAS field, established search algorithms (e.g., NSGA-II or MOEA/D) are frequently adopted without specific enhancements to the search process. As Table 1 demonstrates, AEG-CTFNAS outperforms MedNAS variants that employ conventional search algorithms. Furthermore, our novel visualization method, which leverages Bayesian inference and is integrated into the search process, enables real-time analysis of module contributions—addressing a less explored domain within TFNAS. These innovations critically advance TFNAS development and utility, which will be emphasized in the final manuscript. Future work could broaden comparisons to more TFNAS methods and their search algorithms.
- Relevance to the medical community and medical contribution [R3]: The innovations in search methodology and visualization within AEG-CTFNAS aim to enhance the accessibility, practicality, and safety of TFNAS technology for the medical imaging community. By drastically reducing search times (e.g., on the ACDC dataset, from over 18,000s with NASUNet to under 370s with AEG-CTFNAS, page 6), our work facilitates more rapid real-world application and potentially lowers development costs. Visualizing module (Fig. 3) contributions improves algorithmic interpretability, critical for safety in medical applications. Furthermore, the search space integrates multi-scale feature processing, employing varied pooling operations to effectively capture diverse lesion sizes characteristic of medical images. Further details are in the final manuscript. These contributions collectively accelerate the development of more efficient and effective models, underscoring their direct benefits to the medical community.
- Selecting other test datasets, for example, BTCV [R3]: Acknowledging the value of broader dataset testing, this study prioritized the ACDC and Endovis 2018 datasets for their direct relevance to potential robotic surgery applications, such as surgical tool identification and segmentation. Experiments on these diverse datasets have already confirmed the algorithm’s effectiveness. AEG-CTFNAS’s robust design indicates strong performance potential on additional benchmarks, with experiments on BTCV notably already affirming its advantages over other methods. These findings will be updated in the future paper. Due to rebuttal rules, we are not allowed to add more experiments, results, or analysis to the current manuscript. Nevertheless, we remain very grateful for all comments regarding further explanation and investigation.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A