acceptodds
Under review as a conference paper at ICLR 2027

Structured Subset-Relative Decision Fields for Medical Image Segmentation from Scarce Scribbles

Abstract

Scribble annotations reduce the labeling cost of medical image segmentation, yet when only a few samples are scribble-annotated, models still struggle to learn reliable anatomical semantics from such scarce supervision. Task-specific priors can provide additional inductive guidance beyond direct label supervision, while they often encode higher-order structures, such as region-level organization, semantic relations, and spatial and geometric regularities, which are not readily accommodated by conventional pixel-level supervision. To address this issue, we propose SRDF, an anatomy- and geometry-aware structured representation framework. It formulates multiclass predictions as subset-relative decision fields with locally calibrated distances, yielding differentiable structured objects that jointly encode region semantics and local boundary geometry. Building on SRDF, we further design anatomical-structure and boundary-geometry constraints, allowing anatomical priors and local image evidence to constrain region organization and boundary localization, respectively, at semantic scopes aligned with the corresponding priors, thereby providing structural and geometric guidance for learning under scarce scribble supervision. Experiments on ACDC and Prostate show that the proposed method outperforms existing methods under few-shot scribble supervision.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical image segmentation, Scribble-supervised learning, Semi-supervised learning

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_7PMLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SRDF for medical image segmentation under extremely scarce scribble annotations. SRDF converts multiclass predictions into subset-relative decision fields and uses local distance calibration to represent regional structure and boundary geometry. Based on this representation, the authors design a class adjacency constraint, a scribble-anchored topology reconstruction constraint, and the Subset-Ribbon boundary geometry constraint.

Critical strengths
  1. The paper unifies multiclass predictions, region-level anatomical relations, and local boundary geometry within differentiable decision fields. The method is structured and targeted to the problem setting.

  2. Subset-relative decision fields preserve consistency with the final multiclass prediction while supporting constraints at different levels, including tissue relations, connectivity, and boundary geometry. This provides a certain degree of methodological novelty.

Critical weaknesses
  1. The method highly depends on manually specified anatomical relations, class adjacency, and subset hierarchies. The paper designs different priors for the cardiac and prostate tasks, so the gains may partly come from dataset-specific knowledge. It remains unclear whether the method can be applied to other organs, pathological conditions, or tasks with unstable anatomical relations.

  2. The paper presents the method as a structured learning framework under sparse scribble supervision, but the class adjacency and topology reconstruction constraints are mainly applied to scribble-annotated images. Unlabeled images are primarily used through Mean Teacher, local distance calibration, and boundary constraints. Therefore, it remains unclear whether the main gains truly come from structured utilization of the large amount of unlabeled data.

Questions for authors
  1. Do the anatomical subsets, class adjacency relations, and topology targets need to be manually designed for each dataset? How does the method perform when the prior relations are incomplete, incorrect, or partially removed?

  2. The boundary geometry constraint searches only near the teacher-predicted boundary. When the teacher substantially misses the target or has a large boundary deviation, can this constraint correct the error, or does it mainly refine an already accurate boundary?

Reviewer_Du1COpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces SRDF, a representation that converts multiclass predictions into subset-relative decision fields consistent with the final argmax segmentation. Local distance calibration supports anatomical adjacency constraints, scribble-anchored reconstruction, and an appearance-based boundary objective. Experiments on ACDC and Prostate demonstrate improvements with five scribble-annotated training cases per dataset.

Critical strengths
  • The paper is easy to follow.
Critical weaknesses
  1. Backbone, pretraining, and tuning budgets are not clearly matched across the main baselines. The paper also lacks comparisons against existing anatomical or topological regularizers integrated into the same Mean Teacher backbone. Internal ablations do not resolve these questions.
  2. The experiments are limited. Evaluation covers two MRI datasets and one annotation budget is not enough and the compared papers are not that budget setting. Thus, the reported results confuse readers and make the fair comparison challenging. Backbone experiments report only complete-model performance, without corresponding baselines, so they do not establish consistent gains across architectures.
  3. Table 3 progressively adds objectives but never replaces SRDF while keeping the priors and training procedure fixed. Consequently, the experiments establish the usefulness of the constraint package, rather than the specific benefit of subset-relative decision fields. Comparisons against probability sums or probability log-odds under matched objectives are essential.
  4. Maximum overlap between soft boundary bands can reflect uncertain predictions or thin separating regions, rather than genuine discrete adjacency. Required contact can also be encouraged through a very localized response. This matters because adjacency produces most of the Prostate improvement. Controls using simple class-presence or region regularization are needed to clarify the mechanism.
  5. Downsampling and distance margins may also alter narrow connections. Thus, unreachable foreground may indicate limited propagation rather than disconnected anatomy. Although the paper acknowledges the surrogate nature of reconstruction, sensitivity to convergence, grid resolution, and seed selection remains necessary.
Questions for authors

Authors need to answer the above questions.

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

This paper proposes SRDF, a structured representation framework that incorporates anatomical priors and boundary-geometry constraints to improve medical image segmentation with limited scribble supervision. While the motivation is reasonable and the experimental results show promising improvements, the paper does not sufficiently demonstrate the effectiveness of several proposed components. In addition, the method relies on predefined anatomical relationships, and the limited evaluation does not adequately address its robustness to anatomical variations or its generalizability to other segmentation tasks.

Critical strengths
  1. The paper attempts to solve a meaningful problem by exploring how anatomical knowledge can provide additional supervision for medical image segmentation under scarce scribble annotations. The motivation is reasonable and has practical significance.
  2. The proposed method provides an interesting approach to incorporating anatomical structure and boundary geometry constraints into segmentation learning. The experimental results show promising improvements over existing methods, suggesting the potential value of anatomical priors under limited supervision.
Critical weaknesses
  1. Although anatomical knowledge can provide additional supervision, anatomical structures and their relationships may vary across patients, especially between healthy subjects and patients with pathological conditions. The authors should clarify whether the proposed anatomical priors are sufficiently representative of such variations and how incorrect or incomplete priors may affect segmentation performance.
  2. As shown in Table 3, the adjacency constraint accounts for most of the performance improvement on the Prostate dataset, while the additional contributions of topology reconstruction and Subset-Ribbon are relatively limited. Therefore, the current results do not sufficiently demonstrate the effectiveness of these components. Further experiments are needed to justify their necessity.
  3. The evaluation is limited to only two MRI datasets, ACDC and Prostate, both involving relatively well-defined anatomical structures. It remains unclear whether the proposed method can generalize to other organs, imaging modalities, or more complex anatomical structures. Additional experiments are needed to demonstrate its generalizability.
  4. How does the proposed topology reconstruction ensure that the reconstructed topology is anatomically correct? Because the reconstruction relies on predefined anatomical relationships and predicted segmentation regions, incorrect priors or initial predictions may lead to an erroneous topology. The authors should discuss how the method handles such cases and whether these errors could negatively affect segmentation performance.
  5. Subset-Ribbon relies on teacher predictions to locate the initial boundaries and searches only within a limited neighborhood (Eq. 9). How effective is this strategy when the teacher's predicted boundaries deviate significantly from the ground truth or when image contrast is insufficient to identify reliable boundaries? Further experiments are needed to evaluate its robustness under these conditions.
Questions for authors
  1. The paper does not sufficiently explain why SRDF is necessary for improving segmentation performance. Could the proposed anatomical and boundary constraints be applied directly to probability maps or other existing representations? A controlled comparison is needed to demonstrate SRDF's specific contribution.
  2. As shown in Table 3, the adjacency constraint accounts for most of the performance improvement on Prostate, while the additional contributions of topology reconstruction and Subset-Ribbon are relatively limited. How can the authors further justify the necessity and effectiveness of these components?
  3. Anatomical structures and relationships can vary significantly across patients, especially under pathological conditions. How does the proposed method handle these variations? Are the predefined anatomical priors sufficiently representative, and what happens when these priors are incorrect?
  4. How can the authors ensure that the reconstructed topology is anatomically correct? If the initial predictions or predefined anatomical relationships are incorrect, could topology reconstruction reinforce these errors and negatively affect segmentation performance?
  5. The evaluation is limited to only two MRI datasets with relatively well-defined anatomical structures. How can the authors demonstrate that the proposed method generalizes to other organs, imaging modalities, and more complex anatomical conditions? The current experimental evidence is insufficient to support broad conclusions about its generalizability.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.