acceptodds
Under review as a conference paper at ICLR 2027

Learning Anatomical Segmentation through Reasoning-Augmented Iterative State Transition Networks

Abstract

Accurate left ventricle (LV) segmentation from cardiac magnetic resonance (CMR) images remains challenging owing to anatomical variability, ambiguous boundaries, and large appearance changes across cardiac phases. Existing segmentation networks typically formulate segmentation as a one-pass prediction problem, limiting their ability to progressively correct intermediate errors. To address this limitation, we propose the Reasoning-Augmented Self-correcting Memory Network (RASM), a novel iterative state-transition learning framework that progressively refines the segmentation through iterative residual state transitions. The proposed framework integrates latent anatomical feature encoding, iterative state-transition learning, adaptive anatomical feature refinement, and multi-stage deep supervision within a unified end-to-end optimization framework. By sharing parameters across refinement stages, RASM progressively improves segmentation quality while maintaining efficient inference. Extensive experiments on the public ACDC benchmark demonstrate that the proposed method consistently outperforms representative CNN-based, Transformer-based, and hybrid segmentation models in both segmentation accuracy and boundary delineation. These results demonstrate the effectiveness of iterative state-transition learning as a general paradigm for robust medical image segmentation.

Primary area:
optimization
Keywords:
Left Ventricle, Cardiac Magnetic Resonance, Image Segmentation.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_8ssrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes RASM, an iterative framework for left ventricle segmentation from cardiac MRI. RASM progressively refines the segmentation through parameter-shared state transitions, adaptive anatomical feature refinement, and multi-stage deep supervision. Experiments on the ACDC dataset show improved segmentation and boundary accuracy over the evaluated baselines.

Critical strengths
  1. Clear and practically motivated iterative refinement framework. The paper addresses a meaningful limitation of one-pass segmentation by explicitly refining intermediate predictions through parameter-shared state transitions. The overall design is intuitive and well aligned with the challenges of ambiguous cardiac boundaries.

  2. Consistent empirical improvements with reasonable efficiency. RASM achieves the best results among the evaluated baselines on ACDC across both overlap and boundary metrics. The ablation studies also show consistent gains from the proposed components.

Critical weaknesses
  1. The formulation is not technically convincing. The paper defines the segmentation state as a probability map, e.g., S0=σ(Dϕ(Fc))S^0=\sigma(D_\phi(F_c)), and then updates it as St+1=σ(St+ΔSt)S^{t+1}=\sigma(S^t+\Delta S^t). Under this formulation, even when the predicted residual is zero, i.e., ΔSt=0\Delta S^t=0, the state still changes because St+1=σ(St)≠StS^{t+1}=\sigma(S^t)\neq S^t in general. This seems inconsistent with the claimed residual/self-correcting interpretation, where a zero correction should ideally preserve the current prediction. Relatedly, the stability statement in Section 5.1, which claims that ∥ΔSt∥≤ϵ\|\Delta S^t\|\le \epsilon implies ∥St+1−St∥≤ϵ\|S^{t+1}-S^t\|\le \epsilon, does not directly follow from the stated update rule. The convergence discussion in Section 5.3 is also weak: applying a shared transition operator for a fixed number of TT iterations guarantees termination, but does not by itself establish convergence. This issue directly concerns the mathematical consistency of the proposed iterative mechanism and is therefore a major factor in my negative overall assessment.

  2. The empirical evidence is not yet sufficient to support the claimed novelty and generality of the framework. The experiments are limited to binary LV segmentation on a single dataset, ACDC, with 100 subjects under five-fold cross-validation. This is relatively narrow for a method presented as a general iterative segmentation paradigm. In addition, the comparison set mainly includes U-Net, UNet++, TransUNet, Swin-Unet, MISSFormer, and nnU-Net, while more recent segmentation baselines are not evaluated. The strongest gain over Swin-Unet is also modest in Dice. This limits my confidence that the proposed framework offers a broadly meaningful improvement beyond the specific benchmark considered, which further contributes to my weak-reject assessment.

Questions for authors
  1. Could the authors clarify the intended interpretation of the update St+1=σ(St+ΔSt)S^{t+1}=\sigma(S^t+\Delta S^t), especially since StS^t is already a probability map and a zero residual still changes the state?

  2. What is the key technical distinction between RASM and standard recurrent/iterative refinement methods, and which specific component justifies the terms “reasoning,” “memory,” and “self-correction”?

  3. Can the authors provide stage-wise results for S0,S1,S2,S3S^0, S^1, S^2, S^3 within the same refinement trajectory to directly demonstrate that successive iterations consistently correct previous segmentation errors?

Reviewer_qqCaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the Reasoning-Augmented Self-correcting Memory Network (RASM) for left ventricle segmentation in cardiac MRI. The method combines iterative segmentation-state refinement, adaptive anatomical feature refinement, and multi-stage deep supervision to progressively improve segmentation predictions. On the ACDC dataset, RASM achieves 96.46% Dice and 3.11 mm HD95, outperforming several CNN- and Transformer-based baselines. The main contribution is an end-to-end framework that integrates iterative refinement with parameter sharing to improve segmentation accuracy while maintaining relatively low computational overhead.

Critical strengths
  1. Effective integration of iterative refinement and feature adaptation. The proposed framework combines segmentation-state updates with adaptive anatomical feature refinement in a reasonably straightforward architecture. Sharing parameters across refinement iterations also helps control model complexity. The ablation study shows consistent improvements from the individual components, supporting the effectiveness of the overall design.

  2. Promising segmentation accuracy and computational efficiency. RASM achieves the best reported performance across Dice, IoU, HD95, and ASD on ACDC. The improvements in boundary-related metrics are particularly relevant for cardiac segmentation. The model also maintains competitive inference speed (19.2 ms) and parameter count (38.9M), suggesting that iterative refinement does not necessarily introduce substantial computational overhead.

Critical weaknesses
  1. Limited methodological novelty and insufficient comparison with existing iterative refinement approaches.

The main components, including residual state updates, channel-wise feature attention, parameter sharing, and deep supervision, are well-established techniques. Although the authors describe the framework as reasoning-augmented and self-correcting, it is unclear what fundamentally distinguishes the proposed state-transition mechanism from conventional recurrent segmentation or iterative refinement networks. The experimental comparisons mainly focus on standard encoder-decoder architectures rather than methods that explicitly perform iterative segmentation refinement. This makes it difficult to assess whether the proposed approach provides a meaningful methodological advance beyond combining existing components.

  1. Limited experimental validation and questionable theoretical analysis.

The experiments are conducted exclusively on ACDC, a relatively small cardiac MRI dataset. The paper does not evaluate generalization to other datasets, anatomical structures, or segmentation tasks, despite presenting the method as a general segmentation framework. Furthermore, the theoretical analysis in Section 5 contains a questionable stability claim. Specifically, the stated inequality ∣St+1−St∣≤ϵ|S^{t+1}-S^t|\leq\epsilon does not generally follow from ∣ΔSt∣≤ϵ|\Delta S^t|\leq\epsilon, because the sigmoid operation is applied to the sum of the current probability state and the residual. Even when the residual is zero, the state can change. Similarly, completing a fixed number of iterations does not establish convergence. These issues weaken the theoretical justification for the proposed iterative optimization framework.

Questions for authors
  1. How does RASM compare against existing recurrent or iterative refinement segmentation methods under matched backbone architectures and computational budgets? Such comparisons would help clarify whether the proposed state-transition mechanism offers advantages beyond standard iterative refinement.

  2. Can the authors provide experiments on an additional segmentation dataset or anatomical structure to demonstrate the generalizability of the framework? In particular, it would be valuable to determine whether the improvements persist beyond ACDC.

  3. Can the authors clarify the theoretical stability and convergence claims in Section 5? In particular, how is the stated bound on consecutive segmentation states justified when the update applies a sigmoid to an already normalized segmentation probability map?

Reviewer_zU9NOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a RASM for left ventricle segmentation in cardiac MRI. The method combines iterative state refinement, adaptive anatomical feature refinement, and deep supervision to progressively improve segmentation accuracy. Experiments on the ACDC dataset show improved segmentation performance compared with several conventional segmentation models.

Critical strengths
  1. The paper tries to address a meaningful problem in medical image segmentation. Iterative refinement is a reasonable approach to improving segmentation accuracy, particularly for challenging anatomical boundaries.
  2. The proposed framework integrates iterative state refinement, adaptive feature refinement, and deep supervision into a unified architecture, providing a potentially useful approach to segmentation improvement.
Critical weaknesses
  1. The paper emphasizes reasoning-augmented and self-correcting segmentation. However, the proposed method appears to be a recurrent segmentation framework that refines iteratively. It is unclear what distinguishes the proposed reasoning mechanism from conventional recurrent feature updates. The authors should clarify how reasoning is explicitly achieved and what makes the proposed framework different from existing recurrent segmentation methods.
  2. The theoretical conclusion in Section 5.1 does not hold. The sigmoid operation can change the segmentation state even when Delta S^t=0. This analysis needs to be corrected.
  3. The claimed iterative self-correction mechanism requires more direct experimental validation. The authors should quantify how many incorrectly classified pixels are corrected at each iteration and how many previously correct pixels become incorrect. This would help show whether the model genuinely self-corrects. In addition, the paper discusses adaptive computation, but the current framework uses a fixed number of iterations rather than allocating computation based on regional complexity. This distinction needs further clarification.
  4. The experimental comparisons mainly focus on general segmentation models. Since the main contribution is iterative state refinement, comparisons with existing recurrent and iterative refinement-based segmentation methods are more relevant and necessary to demonstrate the novelty and effectiveness of the proposed approach.
  5. The experimental evaluation is limited to a single dataset and focuses only on LV segmentation. It is unclear whether the proposed framework is specifically designed for this anatomical structure or can generalize to other segmentation tasks. Additional experiments on different anatomical structures and datasets are needed to evaluate its generalizability.
Questions for authors
  1. What distinguishes the proposed reasoning-augmented mechanism from conventional recurrent segmentation? How is reasoning explicitly implemented and validated in the proposed framework?
  2. Can the authors provide quantitative evidence of iterative self-correction, specifically the number of incorrectly classified pixels corrected and previously correct pixels misclassified at each iteration? Also, how does the proposed method achieve adaptive computation when all regions undergo the same fixed number of iterations?
  3. Why are the experimental comparisons limited to general segmentation models? Since iterative refinement is the main contribution, comparisons with existing recurrent and iterative refinement-based methods are necessary to establish its effectiveness and novelty.
  4. Is the proposed framework specifically designed for LV segmentation, or is it applicable to other anatomical structures? Can the authors provide additional experimental evidence to support its generalizability beyond the ACDC dataset?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.