acceptodds
Under review as a conference paper at ICLR 2027

SageProto: Mitigating Class-Wise Learning Disparity in Semi-Supervised Multi-Organ Segmentation

Abstract

Medical image segmentation requires dense expert annotations, motivating semi-supervised learning with limited labeled and abundant unlabeled data. However, anatomical classes differ markedly in learning difficulty, creating pronounced class-wise learning disparity under limited supervision: well-learned classes form reliable visual representations, whereas difficult classes remain under-optimized. In semi-supervised learning, unlabeled supervision and class representation learning remain closely tied to model predictions and visual evidence. These class-wise disparities may therefore persist throughout training. We propose SageProto, a semantic-anchor prototype framework for semi-supervised multi-organ segmentation that introduces textual semantic references into representation learning, prototype estimation, and decoding. SageProto encodes anatomical descriptions into textual semantic anchors that provide class-specific references independent of visual representation strength. These anchors modulate visual representations and initialize class prototypes, which are iteratively rectified with visual evidence and used as class-specific queries for textual and visual retrieval during decoding. Prototype regularization further constrains cross-view consistency, class semantics, and inter-class geometry. By reducing difficult-class dependence on weak visual evidence, SageProto mitigates class-wise learning disparity. Extensive experiments on two multi-organ segmentation benchmarks involving 11 semi-supervised methods show that SageProto improves the best existing Dice by 2.08% on Synapse and 2.31% on FLARE22.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Semi-supervised learning, Medical image segmentation, Prototype learning, Vision-language learning, Class-wise Learning Disparity

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_Rfn8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SageProto, a semantic-anchor prototype framework for semi-supervised multi-organ segmentation. The method aims to mitigate class-wise learning disparity by introducing textual anatomical knowledge into visual representation learning, prototype estimation, and decoding. Specifically, semantic anchors are constructed from anatomical descriptions and used to modulate visual features, initialize and rectify class prototypes, and guide cross-modal feature interaction. Prototype regularization is further introduced to encourage semantic consistency and preserve inter-class relationships. Experiments on Synapse, FLARE22, and AMOS demonstrate improvements over several existing semi-supervised segmentation methods, particularly for some challenging anatomical structures.

Critical strengths
  1. Class-wise performance disparity is particularly important in multi-organ segmentation, where anatomical structures vary substantially in size, appearance, and segmentation difficulty. Introducing anatomical semantic information to support poorly learned classes is a reasonable direction, and the proposed framework provides a systematic way to incorporate such information into prototype learning.
  2. The authors evaluate SageProto on Synapse, FLARE22, and AMOS, with comparisons against several recent semi-supervised segmentation methods. The reported improvements on challenging structures are encouraging. The component ablations and training analyses also provide useful initial evidence for the proposed design.
Critical weaknesses
  1. SageProto combines semantic modulation, iterative prototype rectification, cross-modal decoder interaction, and multiple regularization terms. Although the ablation studies indicate that these components improve performance, it remains unclear whether the improvements primarily arise from the anatomical semantics or from the additional modeling capacity and feature interactions. A more controlled comparison with non-semantic or randomly initialized anchors would help establish the specific value of the proposed semantic guidance. This is important because the central contribution of the paper is the use of semantic anchors to alleviate class-wise learning disparity. 2.There are inconsistencies in the reported experimental results, and the evaluation protocol needs clarification. For example, the Synapse result is 65.85% DSC in Table 1 but 65.29% in Table 2. Similarly, the FLARE22 result is 86.63% in Table 1 but 86.21% in Table 8. Table 8 is labeled as using 10% annotated data, whereas the same DSC value is discussed in connection with the 50% supervision setting. These discrepancies make it difficult to determine which results correspond to the final configuration. Furthermore, the paper does not clearly establish the variability of the reported improvements across different random seeds or labeled-data selections, which is especially important under limited supervision.
Questions for authors
  1. Could the authors explain the inconsistencies in the reported results across Tables 1, 2, and 8?
  2. Could the authors clarify how the experimental results demonstrate that the proposed semantic anchors are responsible for mitigating class-wise learning disparity?
Reviewer_VfDrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SageProto, which constructs semantic anchors from textual anatomical descriptions to mitigate class-wise learning disparity in semi-supervised multi-organ segmentation. The method introduces semantic information into visual feature modulation, prototype estimation, and decoding, and uses prototype regularization to constrain cross-view consistency, semantic consistency, and inter-class relations.

Critical strengths
  1. The paper focuses on the specific problem that difficult organs continue to receive low-quality supervision because of their weak visual representations. Introducing anatomical text as complementary information provides a reasonable approach to this problem.

  2. The semantic anchors are used not only for feature fusion, but also for prototype initialization and inter-class relation constraints, allowing textual knowledge to continuously influence class representations. This gives the method a coherent design.

Critical weaknesses
  1. There are clear inconsistencies between the results in the main experiments and the ablation experiments. On Synapse, the results for Gal., mean Dice, and ASSD differ between Table 1 and the ablation results in Table 2. On FLARE22, the results for most organs differ substantially between Table 1 and Table 8. The authors should check and explain these discrepancies, as they affect the credibility of the organ-wise analysis and the conclusions of the paper.

  2. SageProto is not effective for all organs. For example, its liver performance on Synapse is lower than that of some baselines. The paper mainly emphasizes the mean Dice and the gains on difficult classes, but does not discuss whether this trade-off between organ classes is consistent with the stated objective of the method.

Questions for authors
  1. Why are the results of the full model different across Tables 1, 2, and 8?

  2. How is “reliable class evidence” determined during prototype rectification? When the teacher misses or misclassifies a difficult organ, how does the method prevent an erroneous class centroid from further shifting the prototype?

  3. Equations (8) and (9) update the weights according to directional consistency relative to the visual centroid, but directional consistency does not necessarily mean that a feature is closer to the correct class centroid. Can the authors explain why this update improves the prototypes and compare it with a simple scheme that directly interpolates toward the reliable visual centroid?

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

This paper focuses on the issue of inconsistent learning performance across different organs in semi-supervised multi-organ segmentation. Specifically, that some organs are easier to learn, while smaller organs or those lacking distinct visual features are often difficult to supervise effectively. To address this, the authors propose "SageProto," which incorporates textual anatomical descriptions of organs to provide additional semantic references for different categories. The method is validated on two major datasets: Synapse and FLARE22.

Critical strengths
  1. There are indeed differences in learning difficulty across various organs; in scenarios with limited annotated data, it is often particularly challenging to accurately learn small organs or those with indistinct boundaries. By focusing on the learning challenges associated with these "difficult" categories, the authors have identified a worthwhile research direction.
  2. The authors conducted experiments on the Synapse, FLARE22, and AMOS datasets, compared various semi-supervised learning methods, and performed ablation studies and comparative analyses across different model modules and varying proportions of annotated data. The experimental results demonstrate significant performance improvements for several organs that were previously difficult to segment.
Critical weaknesses
  1. The paper proposes four modules, but since semantic modulation, prototype initialization, and iterative refinement are already well-established concepts, and the combination of consistency losses is also quite common, this work appears more like an engineering-oriented assembly of existing methods.
  2. In Table 1, the average Dice score for "SageProto" on the Synapse dataset is 65.85%, whereas the score for the complete model in Table 2 is 65.29%; for the Gallbladder class, the figures are 38.36% and 33.86%, respectively. Similarly, for the FLARE22 dataset, Table 1 and Table 8 report 86.63% and 86.21%, respectively. I find these data points confusing.
Questions for authors

Please explain the inconsistency in this data.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.