acceptodds
Under review as a conference paper at ICLR 2027

mCNA: Multi-Clone Communication Neural Architecture Search for 3D Medical Image Segmentation

Abstract

Volumetric medical image segmentation requires precise delineation of anatomical structures under substantial variations in organ size, shape, and appearance, while being constrained by high memory and computational costs of 3D processing. Neural Architecture Search (NAS) provides a principled approach for automated architecture design, but directly searching large 3D segmentation networks is prohibitively expensive due to the combinatorial complexity of jointly optimizing cell operations, skip connections, network topology, and model capacity under dense volumetric representations. To address this challenge, we propose Multi-Clone Communication Neural Architecture (mCNA), a search-efficient 3D NAS framework that decomposes 3D architecture optimization into compact cell search and clonable macro-level expansion. Instead of searching a complete large-scale U-Net topology, mCNA first discovers an optimized 3D U-Net backbone and then constructs a larger architecture by replicating the searched unit with lightweight Communication Units for inter-clone feature exchange and cooperative refinement. To avoid costly discrete skip-topology exploration, we introduce a Path Attention Module that performs differentiable adaptive fusion of multi-scale encoder features. Furthermore, we incorporate structure-preserving wavelet-based downsampling operations with pre-optimized and fixed filters into the search space to retain fine-grained volumetric information during downsampling. Experiments on Synapse and the Medical Segmentation Decathlon benchmark demonstrate that mCNA achieves the best overall performance among the evaluated methods under aligned preprocessing, training, and evaluation protocols. Extensive ablation studies verify the effectiveness of the proposed cell search, clone communication, adaptive skip fusion, and wavelet-based downsampling components. The searched architecture can also be transferred across datasets without additional architecture search, demonstrating improved search efficiency and architectural generalizability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Medical Image Segmentation, Neural Architecture Search, Differentiable Architecture Search, Multi-Scale Feature Fusion, Wavelet Pooling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UBVqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a neural architecture search framework for 3D medical image segmentation that separates local feature extraction from global context aggregation. The central design introduces two learned components: a searched local cell that combines wavelet transforms, convolutions, and channel weighting operations within a differentiable search space, and a Communication Unit (CU) that exchanges information across multiple parallel network clones through a gated cross-clone messaging mechanism. The motivation is that existing segmentation networks couple local and global operations within the same architectural units, making it difficult to independently optimize each. By decoupling these roles and searching for efficient local cells, the framework aims to reduce the parameter count and FLOPs of the search supernet while maintaining competitive segmentation accuracy.

The framework is evaluated on multiple 3D medical segmentation benchmarks including Synapse multi-organ segmentation, and supplementary 2D experiments are also provided. Ablations examine the contribution of the CU, the searched versus hand-designed cells, clone count, network depth, wavelet operations, and computational cost. The appendix provides detailed analyses of communication patterns, clone scaling, depth sensitivity, wavelet component effects, and efficiency comparisons.

Critical strengths

S1: The decoupled design of local search and global expansion has practical value. Separating the searched local cell from the cross-clone communication module provides a cleaner architectural decomposition than methods that couple these roles. The reported reductions in supernet parameter count and FLOPs relative to comparable search frameworks are meaningful from an efficiency standpoint.

S2: Experiments cover multiple 3D segmentation tasks with 2D supplementary results. The evaluation is not limited to a single benchmark, and the inclusion of both 3D and 2D settings provides broader coverage than is typical for architecture search papers in medical imaging.

S3: The appendix provides relatively rich supporting analyses. Analyses of communication patterns, clone count scaling, network depth, wavelet component contributions, and computational cost are included, which gives reviewers more material to assess the design choices than the main paper alone would allow.

Critical weaknesses

W1: The architectural advantage of multi-clone expansion is not sufficiently isolated. The paper motivates multi-clone expansion as a way to achieve a favorable accuracy–cost tradeoff. However, the clone-number ablation changes model capacity and computational cost together with the number of clones. It therefore does not establish whether the gains arise specifically from the multi-clone organization or from increased capacity. A comparison with a wider single U-Net using the same searched cells and PAM, under matched parameter or computational budgets, would help isolate the benefit of clone-based expansion. The existing results demonstrate the effectiveness of the overall system, but provide limited evidence that this macro-architectural choice offers an advantage over alternative ways of allocating the same budget.

W2: The benefit of gated communication remains confounded with additional model capacity. In Table 12, replacing No Comm. with CU increases the parameter count from 26.92M to 42.86M, approximately 59%. Direct addition improves performance with almost no parameter increase, supporting the usefulness of cross-clone communication. However, the additional improvement achieved by CU cannot be cleanly attributed to its gating mechanism, since CU also introduces substantial convolutional capacity. Although the authors acknowledge this limitation, two controls would strengthen the attribution: retaining all CU convolutions while blocking cross-clone messages, and comparing against a parameter-matched non-gated fusion module. These comparisons would help distinguish the effects of communication, gating, and additional feature transformation.

W3: The advantage and stability of the architecture search require stronger evidence. Table 8 shows that replacing hand-designed cells with searched cells improves Dice by 0.69 percentage points without CU and PAM, and by 2.25 percentage points with both modules. These results support the usefulness of the discovered architecture. However, the hand-designed baseline relies on standard convolutions and average pooling and does not establish whether the search procedure outperforms simpler alternatives within the same operation space. Budget-controlled comparisons with randomly sampled architectures and stronger manually configured cells would help distinguish the benefit of architecture search from that of the search space itself. In addition, the paper does not clearly establish whether the reported three runs include independent architecture searches or only retraining of a fixed searched architecture. Reporting variability across independent searches would strengthen the evidence for the robustness of the search procedure.

Questions for authors

The independence between architecture selection and final evaluation is unclear. Appendix D.1 states that network depth, node count, and clone count are chosen based on validation Dice and efficiency, and Appendices E.4–E.6 present these comparisons on Synapse. The main Synapse results use 18 training cases and 12 validation cases. It is therefore unclear whether the 12 validation cases used to select clone count, depth, and node count are the same cases used to report the main results. If they are, the reported numbers serve simultaneously as model selection criteria and final evaluation, which introduces optimistic bias from selection. Three training runs with different random seeds do not resolve this issue. The authors should clarify the data usage at each stage, confirm whether an independent held-out set or nested validation was used, and state whether baseline methods received a comparable hyperparameter tuning budget.

Reviewer_nrnoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes neural architecture search methods mCNA for 3D medical image segmentation. The framework decomposes 3D architecture search into cell search and macro level expansion. Simply speaking, the mCNA is a cell-level search on small unet search space (they also add a hand-crafted attention module to the unit) -> replicate those U-Nets and connect with communication unit-> fuse the results. They validated the network on Synapse abdomen CT, MSD, and 2D experiments on CHAOS, BAGLS, KVASIR and showed they get good results.

Critical strengths

The overall paper is clear and authors did quite a lot experiments covering both 3D and 2D and enough baseline architectures. Also the author proposed to stack the searched unet to save search complexity for the macro-level.

Critical weaknesses
  1. The contribution is mainly about medical segmentation network architecture, instead of contribution to NAS. Only cell-search is NAS based and the rest of the model are mostly hand-crafted. Ablation showed some evidence those elements helped, but not convincing enough.
  2. Since from NAS perspective, the novelty is minimum, so I look the paper as a architecture paper for medical image segmentation. However, I don't see anything novel in the architecture design principles, the design seems very ad-hoc without solid proof / novel intuition, at least not enough for ICLR standards.
  3. If the novelty is limited, I look at the performance. Although the result table compared with different baselines, whether those baselines are well-trained and is under fair comparison is unknown. Especially look at the MSD results, those numbers are substantially lower than the MSD leaderboard and numbers reported by the baseline papers. A fair comparison on MSD test leaderboard results is more convincing. Synapse is such a saturated small dataset and I don't trust it numbers at all. For an architecture paper, prove it either in fair-comparison challenges or large/diverse datasets.
Questions for authors

What is the central methodological novelty over existing cell-level NAS?

Reviewer_TMp9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes mCNA, a neural architecture search (NAS) framework for 3D medical image segmentation. To reduce the computational complexity of searching large 3D architectures, the proposed method performs cell-level architecture search within a single U-Net backbone and subsequently expands the searched architecture into multiple communicating U-Net clones. A gated Communication Unit (CU) facilitates feature exchange between adjacent clones, while a Path Attention Module (PAM) adaptively aggregates multi-scale encoder features without explicitly searching skip-connection topologies. Additionally, the authors introduce orthogonality- and reversibility-constrained learnable discrete wavelet pooling (Orth. LDW) as candidate downsampling operations to preserve structural information.

Experiments on Synapse and all ten tasks of the Medical Segmentation Decathlon (MSD) demonstrate competitive segmentation performance, outperforming the evaluated baselines in terms of overall Dice scores. The paper also includes component-wise ablation studies and computational analyses to evaluate the proposed framework.

Critical strengths
  1. The paper proposes an intuitive and computationally efficient NAS framework by separating cell-level architecture search from macro-level network expansion.

  2. The proposed method achieves strong segmentation performance across Synapse and MSD benchmarks, outperforming the evaluated baselines in most tasks.

  3. The paper provides extensive ablation studies examining the contributions of individual components, communication strategies, and computational efficiency.

Critical weaknesses
  1. Limited NAS comparisons and unclear contribution of architecture search. Although NAS is a central contribution, DiNTS is the only NAS-based baseline. Furthermore, the ablation results suggest that the performance improvement from NAS itself is relatively modest compared to other components, particularly PAM. The necessity of NAS within the proposed framework is therefore not sufficiently demonstrated.

  2. Insufficient capacity-matched comparisons. The multi-clone architecture increases the number of parameters and computational cost. Without comparisons against larger single-network architectures under similar computational budgets, it is difficult to determine whether the improvements result from the proposed communication mechanism or increased model capacity.

  3. Concerns about baseline evaluation fairness. All competing methods are retrained under a unified protocol, which may not preserve their original optimized configurations. Moreover, evaluation under the established benchmark protocols of prior methods is limited, making it difficult to assess whether the reported improvements generalize beyond the authors' experimental settings.

  4. Limited reproducibility and statistical validation. It is unclear whether architecture search was repeated across different random seeds. Additionally, the MSD experiments and most ablation studies rely on single-run results, despite relatively small improvements on some tasks.

Questions for authors
  1. Could the authors provide additional NAS-based comparisons and clarify whether a manually designed U-Net equipped with the same CU, PAM, and Orth. LDW could achieve comparable performance without architecture search?

  2. Could the authors compare the multi-clone architecture against a capacity-matched single U-Net to isolate the effectiveness of the proposed communication mechanism?

  3. Were the original training configurations of competing methods preserved or appropriately tuned? Could the authors additionally evaluate mCNA under representative established benchmark protocols to enable more direct comparisons with previously reported results?

  4. Was architecture search repeated with different random seeds, and how consistent were the resulting architectures and segmentation performance?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.