pith. machine review for the scientific record. sign in

arxiv: 2605.08238 · v1 · submitted 2026-05-07 · 💻 cs.CV · cs.AI· cs.ET· cs.LG

Recognition: 2 theorem links

· Lean Theorem

Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation

Farhana Yasmin , Mahade Hasan , Haipeng Liu , Amjad Ali , Ghulam Muhammad , Yu Xue
Authors on Pith no claims yet

Pith reviewed 2026-05-12 01:22 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.ETcs.LG
keywords cardiac MRI segmentationneural architecture searchevolutionary algorithmUNetresource-aware optimizationDice similarity coefficientHausdorff distanceACDC dataset
0
0 comments X

The pith

CardiacNAS discovers a UNet variant with 93.22% DSC using only 3.58M parameters and 14.56 GFLOPs for cardiac MRI segmentation.

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

The paper presents an evolutionary neural architecture search framework called CardiacNAS that builds on a UNet-like supernet and a cardiac-specific search space covering depth, width, kernels, attention, fusion, activations, dropout, and residual scaling. The method treats segmentation accuracy, measured by Dice similarity coefficient and 95th-percentile Hausdorff distance, as objectives to be maximized while minimizing model size and floating-point operations under fixed compute budgets. Candidate networks are sampled, trained briefly on proxy budgets, and evolved via crossover, mutation, and elitist selection. On the ACDC dataset the final architecture outperforms six hand-designed baselines in the accuracy-efficiency tradeoff, reaching 93.22 percent average DSC and 4.73 mm HD95. This matters for clinical workflows because accurate yet lightweight segmentation enables quantitative ventricular assessment on devices with limited memory or processing power.

Core claim

CardiacNAS couples a UNet-like supernet with a cardiac-aware search space and performs resource-aware evolutionary search that jointly optimizes DSC and HD95 against parameter count and FLOPs. Architectures are instantiated, trained under proxy budgets, and refined through crossover, mutation, and elitist selection. The best resulting model attains 93.22 percent average DSC and 4.73 mm HD95 with 3.58 million parameters and 14.56 GFLOPs, exceeding the accuracy-efficiency balance of six state-of-the-art hand-designed methods. Analyses show that the searched attention and fusion choices together with residual scaling improve boundary fidelity and training stability.

What carries the argument

The resource-aware evolutionary optimizer that evolves candidate architectures from a UNet-like supernet under a cardiac-aware search space, using proxy-budget training to rank designs by the joint objectives of DSC, HD95, model size, and FLOPs.

If this is right

  • Attention and fusion modules selected by the search improve boundary delineation in low-contrast cardiac regions.
  • Residual scaling in the discovered architectures contributes to more stable training and consistent performance across scans.
  • Explicit reporting of parameter counts and FLOPs allows direct assessment of deployability on clinical hardware.
  • The framework supplies a repeatable procedure for generating efficient CMR segmentation models under explicit compute constraints.

Where Pith is reading between the lines

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

  • The same search strategy could be transferred to other medical imaging tasks such as CT or ultrasound cardiac analysis with only modest redefinition of the search space.
  • Repeating the evolutionary search several times and comparing the resulting top models would indicate how sensitive final performance is to random seed and population initialization.
  • Correlation studies between design factors and performance could be extended to identify reusable building blocks for efficient medical segmentation networks beyond the ACDC dataset.

Load-bearing premise

Proxy-budget training during the search produces architecture rankings that remain valid after full convergence, and the cardiac-aware search space plus evolutionary operators are sufficient to find designs meaningfully better than the six hand-designed baselines.

What would settle it

Retrain the top-searched architecture and the six baselines to full convergence on the ACDC dataset and compare the final DSC and HD95 values; if the searched model no longer leads in the accuracy-efficiency tradeoff the central claim fails.

Figures

Figures reproduced from arXiv: 2605.08238 by Amjad Ali, Farhana Yasmin, Ghulam Muhammad, Haipeng Liu, Mahade Hasan, Yu Xue.

Figure 1
Figure 1. Figure 1: Overall framework of CardiacNAS, showing supernet initialization, candidate generation through crossover and mutation, and evolutionary selection [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Correlation matrix across 200 architectures, showing how attention [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on ACDC cardiac MRI. Columns: image, GT overlay, and predictions from U-Net, nnU-Net, TransUNet, BiX-NAS, HCT-Net, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: ACDC validation curves of average HD95 (mm) for the same seven [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
read the original abstract

Cardiac magnetic resonance (CMR) segmentation underpins quantitative assessment of ventricular structure and function, yet reliable delineation remains difficult due to low tissue contrast, fuzzy boundaries, and inter scan variability. We present CardiacNAS, an evolutionary neural architecture search (NAS) framework that couples a UNet like supernet with a cardiac aware search space spanning depth width, kernel size, filter size, attention, fusion, activation, dropout, and residual scaling. The search is explicitly resource aware, jointly optimizing dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95) versus model size and floating point operations (FLOPs) under fixed compute budgets. Candidate architectures are instantiated from the supernet, trained with proxy budgets, and evolved through crossover, mutation, and elitist selection. We evaluate on the ACDC dataset and compare against six state of the art methods, using qualitative comparisons, learning curve analyses, and design factor correlation studies. The resulting model attains 93.22% average DSC and 4.73 mm HD95 with 3.58M parameters and 14.56 GFLOPs, demonstrating a favorable accuracy efficiency trade off. Analyses indicate that searched attention and fusion choices, together with residual scaling, contribute to improved boundary fidelity and stability. CardiacNAS offers a principled, resource aware approach to deployable CMR segmentation with transparent reporting of architectural complexity and compute budgets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper introduces CardiacNAS, an evolutionary neural architecture search (NAS) framework for cardiac MRI segmentation. It couples a UNet-like supernet with a cardiac-aware search space spanning depth, width, kernel size, filter size, attention, fusion, activation, dropout, and residual scaling. The search is resource-aware, jointly optimizing DSC and HD95 against model size and FLOPs under fixed budgets. Candidates are instantiated from the supernet, trained with proxy budgets, and evolved via crossover, mutation, and elitist selection. On the ACDC dataset, the resulting model is compared to six state-of-the-art hand-designed methods using quantitative metrics, qualitative results, learning curves, and design-factor correlations. The discovered architecture attains 93.22% average DSC and 4.73 mm HD95 with 3.58M parameters and 14.56 GFLOPs.

Significance. If the central empirical claims hold, the work offers a practical, transparent method for discovering efficient architectures tailored to CMR segmentation. The explicit resource awareness, domain-specific search space, and post-search analyses of attention/fusion/residual choices provide useful insights into factors affecting boundary fidelity and stability, supporting deployable models in clinical settings.

major comments (1)
  1. [Methods (search procedure)] Methods (search procedure): The evolutionary search ranks and selects architectures exclusively under proxy training budgets, with full-budget training applied only to the single final model whose performance (93.22% DSC, 4.73 mm HD95) is reported. No rank-correlation study, ablation, or validation is described showing that proxy rankings predict full-budget ordering. This directly affects the claim that the cardiac-aware space plus evolutionary operators discovered a meaningfully superior accuracy-efficiency tradeoff versus the six baselines, because any material rank reversal under full training would mean a different architecture should have been selected.
minor comments (1)
  1. [Abstract] Abstract: 'state of the art methods' should read 'state-of-the-art methods'.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comment on the search procedure below and will revise the paper to incorporate additional validation as suggested.

read point-by-point responses

  1. Referee: The evolutionary search ranks and selects architectures exclusively under proxy training budgets, with full-budget training applied only to the single final model whose performance (93.22% DSC, 4.73 mm HD95) is reported. No rank-correlation study, ablation, or validation is described showing that proxy rankings predict full-budget ordering. This directly affects the claim that the cardiac-aware space plus evolutionary operators discovered a meaningfully superior accuracy-efficiency tradeoff versus the six baselines, because any material rank reversal under full training would mean a different architecture should have been selected.

    Authors: We acknowledge that the current manuscript does not include an explicit rank-correlation analysis between proxy and full-budget training. Proxy budgets are used during search to manage computational cost, which is standard in NAS literature, but we agree that demonstrating their predictive power would strengthen the claims. In the revised version, we will add a dedicated analysis: we will re-train a representative subset of 12 architectures (selected across the evolutionary generations) under full budgets and compute Spearman's rank correlation on DSC and HD95. We will report the correlation coefficients, discuss any rank reversals, and clarify that all six baseline methods were trained and evaluated with full budgets for fair comparison of the final model. This addition will be placed in the Methods or Experiments section with supporting tables. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical NAS results are direct experimental outcomes

full rationale

The paper describes an evolutionary NAS procedure that instantiates candidates from a supernet, trains them under proxy budgets, applies crossover/mutation/elitist selection, and reports the fully-trained performance of the selected architecture on the ACDC dataset. The headline numbers (93.22% DSC, 4.73 mm HD95, 3.58 M params, 14.56 GFLOPs) are measured outputs of this process, not quantities that reduce by the paper's own equations or self-citations to the search inputs. No self-definitional loops, fitted-input predictions, or load-bearing self-citations appear in the provided text; the cardiac-aware search space and resource-aware objective are stated as design choices rather than derived necessities. The result is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Insufficient information in the abstract to enumerate specific free parameters, axioms, or invented entities; the work relies on standard assumptions of supervised segmentation (i.i.d. data, Dice as proxy for clinical utility) and evolutionary search convergence.

pith-pipeline@v0.9.0 · 5570 in / 1130 out tokens · 25072 ms · 2026-05-12T01:22:31.717130+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages

  1. [1]

    MECardNet: A novel multi-scale convolutional ensemble model with adaptive deep supervision for precise cardiac MRI segmen- tation,

    H. Aghapanah, R. Rasti, F. Tabesh, H. Pouraliakbar, H. Sanei, and S. Kermani, “MECardNet: A novel multi-scale convolutional ensemble model with adaptive deep supervision for precise cardiac MRI segmen- tation,”Biomedical Signal Processing and Control, vol. 100, p. 106919, 2025

    work page 2025

  2. [2]

    URCA: Uncertainty-based region clipping algorithm for semi-supervised medical image segmentation,

    C. Qin, Y . Wang, and J. Zhang, “URCA: Uncertainty-based region clipping algorithm for semi-supervised medical image segmentation,” Computer Methods and Programs in Biomedicine, vol. 254, p. 108278, 2024

    work page 2024

  3. [3]

    Exploring inherent consistency for semi-supervised anatomical structure segmentation in medical imaging,

    W. Huang, L. Zhang, Z. Wang, and L. Wang, “Exploring inherent consistency for semi-supervised anatomical structure segmentation in medical imaging,”IEEE Transactions on Medical Imaging, vol. 43, no. 11, pp. 3731–3741, 2024

    work page 2024

  4. [4]

    DCA-Net: Data-driven collaborative assistance network for semi-supervised medical segmentation,

    Y . Chen, C. Wang, and B. Zhao, “DCA-Net: Data-driven collaborative assistance network for semi-supervised medical segmentation,” in2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp. 1430–1437, 2024

    work page 2024

  5. [5]

    ScribFormer: Transformer makes CNN work better for scribble-based medical image segmentation,

    Z. Li, Y . Zheng, D. Shan, S. Yang, Q. Li, B. Wang, Y . Zhang, Q. Hong, and D. Shen, “ScribFormer: Transformer makes CNN work better for scribble-based medical image segmentation,”IEEE Transactions on Medical Imaging, vol. 43, no. 6, pp. 2254–2265, 2024

    work page 2024

  6. [6]

    Evo-GrayNet: Colon polyp detection and segmentation using evolutionary network architecture search,

    F. Yasmin, Y . Xue, M. Hasan, M. Gabbouj, M. K. Hasan, K. Au- rangzeb, and G. Muhammad, “Evo-GrayNet: Colon polyp detection and segmentation using evolutionary network architecture search,”IEEE Transactions on Instrumentation and Measurement, pp. 1–16, 2025

    work page 2025

  7. [7]

    Hresformer: Hybrid residual transformer for vol- umetric medical image segmentation,

    S. Ren and X. Li, “Hresformer: Hybrid residual transformer for vol- umetric medical image segmentation,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 6, pp. 10558–10566, 2025

    work page 2025

  8. [8]

    Rule mining of early diabetes symptom and applied supervised machine learning and cross validation approaches based on the most important features to predict early-stage diabetes,

    M. Hasan, F. Yasmin, and L. Deng, “Rule mining of early diabetes symptom and applied supervised machine learning and cross validation approaches based on the most important features to predict early-stage diabetes,”IJIRMPS-International Journal of Innovative Research in Engineering & Multidisciplinary Physical Sciences, vol. 11, no. 3, pp. 1– 31, 2023

    work page 2023

  9. [9]

    Hidiff: Hybrid dif- fusion framework for medical image segmentation,

    T. Chen, C. Wang, Z. Chen, Y . Lei, and H. Shan, “Hidiff: Hybrid dif- fusion framework for medical image segmentation,”IEEE Transactions on Medical Imaging, vol. 43, no. 10, pp. 3570–3583, 2024

    work page 2024

  10. [10]

    Developing advanced AI models with fusion data,

    F. Yasmin, M. Hasan, and Y . Xue, “Developing advanced AI models with fusion data,” inFeature Fusion for Next-Generation AI: Building Intelligent Solutions from Medical Data, pp. 181–193, Springer Nature Switzerland Cham, 2025

    work page 2025

  11. [11]

    Learnable prompting sam-induced knowledge distillation for semi- supervised medical image segmentation,

    K. Huang, T. Zhou, H. Fu, Y . Zhang, Y . Zhou, C. Gong, and D. Liang, “Learnable prompting sam-induced knowledge distillation for semi- supervised medical image segmentation,”IEEE Transactions on Medical Imaging, vol. 44, no. 5, pp. 2295–2306, 2025

    work page 2025

  12. [12]

    Merging context clus- tering with visual state space models for medical image segmentation,

    Y . Zhu, D. Zhang, Y . Lin, Y . Feng, and J. Tang, “Merging context clus- tering with visual state space models for medical image segmentation,” IEEE Transactions on Medical Imaging, vol. 44, no. 5, pp. 2131–2142, 2025

    work page 2025

  13. [13]

    Swin-umamba†: Adapting mamba-based vision foundation models for medical image segmentation,

    J. Liu, H. Yang, H.-Y . Zhou, L. Yu, Y . Liang, Y . Yu, S. Zhang, H. Zheng, and S. Wang, “Swin-umamba†: Adapting mamba-based vision foundation models for medical image segmentation,”IEEE Transactions on Medical Imaging, pp. 1–1, 2024

    work page 2024

  14. [14]

    Advances in brain imaging technologies: A comprehensive overview,

    M. Hasan, F. Yasmin, X. Yu, and A. Karim, “Advances in brain imaging technologies: A comprehensive overview,”Brain Networks in Neuroscience: Personalization Unveiled Via Artificial Intelligence, vol. 1, no. 1, pp. 11–40, 2025

    work page 2025

  15. [15]

    EPSO-Net: A multi-objective evolutionary neural architecture search with PSO-guided mutation fusion for explainable brain tumor segmentation,

    F. Yasmin, Y . Xue, M. Hasan, and G. Muhammad, “EPSO-Net: A multi-objective evolutionary neural architecture search with PSO-guided mutation fusion for explainable brain tumor segmentation,”Information Fusion, p. 104119, 2026

    work page 2026

  16. [16]

    Mde-evonas: Automatic network architecture design for monocular depth estimation via evolutionary neural architecture search,

    Z. Yu, H. Zhang, R. Liu, S. Dai, X. Chen, W. Sheng, and Y . Jin, “Mde-evonas: Automatic network architecture design for monocular depth estimation via evolutionary neural architecture search,”Swarm and Evolutionary Computation, vol. 93, p. 101837, 2025

    work page 2025

  17. [17]

    Psnas-net: Hybrid gradient-physical optimizationfor efficient neural architecture search in customized medical imaging analysis,

    Z. Zechen, H. Xuelei, Z. Fengjun, and H. Xiaowei, “Psnas-net: Hybrid gradient-physical optimizationfor efficient neural architecture search in customized medical imaging analysis,”Expert Systems with Applica- tions, vol. 288, p. 128155, 2025

    work page 2025

  18. [18]

    Mixed-GGNAS: Mixed Search-space NAS based on genetic algorithm combined with gradient descent for medical image segmen- tation,

    M. Hu, J. Li, Y . Dong, Z. Zhang, W. Liu, P. Zhang, Y . Ping, L. Jiang, and Z. Yu, “Mixed-GGNAS: Mixed Search-space NAS based on genetic algorithm combined with gradient descent for medical image segmen- tation,”Expert Systems with Applications, vol. 289, p. 128338, 2025

    work page 2025

  19. [19]

    U-Net: Convolutional Net- works for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Net- works for Biomedical Image Segmentation,” inMedical Image Comput- ing and Computer-Assisted Intervention (MICCAI), vol. 9351 ofLecture Notes in Computer Science, pp. 234–241, Springer, 2015

    work page 2015

  20. [20]

    Ftrans- deeplab: Multimodal fusion transformer-based deeplabv3+ for remote sensing semantic segmentation,

    H. Feng, Q. Hu, P. Zhao, S. Wang, M. Ai, D. Zheng, and T. Liu, “Ftrans- deeplab: Multimodal fusion transformer-based deeplabv3+ for remote sensing semantic segmentation,”IEEE Transactions on Geoscience and Remote Sensing, vol. 63, no. 1, pp. 1–18, 2025

    work page 2025

  21. [21]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

    F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,”Nature methods, vol. 18, no. 2, pp. 203–211, 2021

    work page 2021

  22. [22]

    Transunet: Rethinking the u-net architecture design for medical image segmentation through the lens of transformers,

    J. Chen, J. Mei, X. Li, Y . Lu, Q. Yu, Q. Wei, X. Luo, Y . Xie, E. Adeli, Y . Wang, M. P. Lungren, S. Zhang, L. Xing, L. Lu, A. Yuille, and Y . Zhou, “Transunet: Rethinking the u-net architecture design for medical image segmentation through the lens of transformers,”Medical Image Analysis, vol. 97, p. 103280, 2024

    work page 2024

  23. [23]

    Multi-objective evolutionary optimization boosted deep neural networks for few-shot medical segmentation with noisy labels,

    H. Li, Y . Zhang, and Q. Zuo, “Multi-objective evolutionary optimization boosted deep neural networks for few-shot medical segmentation with noisy labels,”IEEE Journal of Biomedical and Health Informatics, vol. 29, no. 6, pp. 4362–4373, 2025

    work page 2025

  24. [24]

    BiX-NAS: searching efficient Bi-directional architecture for medical image segmentation,

    X. Wang, T. Xiang, C. Zhang, Y . Song, D. Liu, H. Huang, and W. Cai, “BiX-NAS: searching efficient Bi-directional architecture for medical image segmentation,” inMedical Image Computing and Computer Assisted Intervention-MICCAI 2021, pp. 229–238, Springer International Publishing, 2021

    work page 2021

  25. [25]

    Hct-net: hybrid CNN-transformer model based on a neural architecture search network for medical image segmentation,

    Z. Yu, F. Lee, and Q. Chen, “Hct-net: hybrid CNN-transformer model based on a neural architecture search network for medical image segmentation,”Applied Intelligence, vol. 53, no. 17, pp. 19990–20006, 2023

    work page 2023

  26. [26]

    Evolutionary neural architecture search for automated MDD diagnosis using multimodal MRI imaging,

    T. Li, N. Hou, J. Yu, Z. Zhao, Q. Sun, M. Chen, Z. Yao, S. Ma, J. Zhou, and B. Hu, “Evolutionary neural architecture search for automated MDD diagnosis using multimodal MRI imaging,”iScience, vol. 27, no. 10, p. 111020, 2024

    work page 2024

  27. [27]

    Transgraphnet: A novel network for medical image segmentation based on transformer and graph convolution,

    J. Zhang, Z. Ye, M. Chen, J. Yu, and Y . Cheng, “Transgraphnet: A novel network for medical image segmentation based on transformer and graph convolution,”Biomedical Signal Processing and Control, vol. 104, p. 107510, 2025

    work page 2025