acceptodds
Withdrawn from ICLR 2027

Reci-Q: A Class-aware Patch Sampling Curriculum for Imbalanced 3D Medical Image Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Class imbalance is a fundamental challenge in medical image segmentation, where sampling strategies have been widely explored to improve the representation of underrepresented classes during training. However, sampling strategies developed for long-tailed classification cannot be directly transferred to segmentation, as dense prediction additionally requires spatial localization to determine where training crops are extracted. Existing sampling strategies for segmentation typically address this issue at a coarse foreground–background level, overlooking the substantially different prevalence of individual foreground classes. More recent adaptive and curriculum-based approaches dynamically adjust sampling during training, but often rely on auxiliary models, online difficulty signals, or additional mechanisms that increase training complexity. To address these limitations, we propose Reci-Q, a dynamic class-aware sampling strategy that explicitly decouples class selection from spatial localization. The sampling distribution is controlled by a single parameter and follows a reciprocal curriculum, transitioning from class-balanced sampling to the natural prevalence distribution and then back to class-balanced sampling. This design provides class-specific rebalancing while fully exploiting the information from prevalent samples, without any auxiliary models or explicit loss redesign. Experiments across multiple 3D medical image segmentation tasks demonstrate that Reci-Q consistently improves segmentation performance, with particularly notable gains for tail classes.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Medical Image Segmentation, Data Sampling Strategy

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

This paper proposes Reci-Q, a class-aware patch sampling curriculum for imbalanced 3D medical image segmentation. The method decouples class selection from spatial patch localization and introduces a reciprocal sampling schedule that transitions from class-balanced sampling to natural prevalence-based sampling and back. Experiments on five Medical Segmentation Decathlon tasks demonstrate improvements over the nnU-Net baseline, particularly for underrepresented classes, without additional models or loss modifications.

Critical strengths
  1. The proposed method provides a lightweight, class-aware sampling strategy that can be easily integrated into existing segmentation frameworks without modifying network architectures, introducing auxiliary models, or increasing training complexity significantly. The reciprocal curriculum is intuitive and straightforward to implement.

  2. The method is evaluated on five 3D medical segmentation tasks with different class imbalance characteristics. The experiments show improvements in both overall and tail-class Dice scores over the nnU-Net baseline. The ablation studies comparing static and dynamic sampling schedules also provide useful evidence for the effectiveness of the proposed curriculum.

Critical weaknesses
  1. The method primarily combines existing class-aware sampling with a predefined curriculum. The choice of the reciprocal schedule lacks sufficient theoretical justification, and the implicit loss-reweighting interpretation is not rigorously established. Thus, the novelty is limited.

  2. The comparisons with existing sampling methods are limited, and the reported improvements are relatively modest. The main experiment only compare the proposed method with four nnU-net based methods.

Questions for authors
  1. The first question is about the motivation of the specific reciprocal schedule (q:0\rightarrow1\rightarrow0)? Could the authors provide a more rigorous justification for why returning to class-balanced sampling improves generalization compared with other schedules, such as cosine or stepwise scheduling?
  2. In Eq. (5), background sampling selects anchor voxels uniformly from the entire volume, meaning that the actual anchor class may differ from the selected class. How does this affect the approximation in Eq. (9) and the claimed relationship between sampling probabilities and class-specific loss contributions?
  3. Could the authors include stronger class-aware patch sampling or adaptive curriculum baselines to better demonstrate the advantages of Reci-Q over existing sampling strategies?
  4. Since the experiments do not use a separate validation set, how were the curriculum schedule and phase ratio selected? Were these choices determined independently of the test sets?
Reviewer_4b89OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a patch-sampling strategy to mitigate class imbalance (specifically background-foreground imbalance) problem in 3D medical segmentation training. A two-stage sampling with a hyperparameter q decides to run class-balanced sampling (q=0) or natural prevalence-based sampling (q=1). The paper suggests q: 0 -> 1 -> 0 curriculum shows good performance by removing background noise in the middle phase and refocusing on the foreground in the last phase. The evaluation shows the best curriculum strategy depends on the dataset type.

Critical strengths

The authors provide a simple strategy to mitigate the foreground-background imbalance problem in medical image segmentation task.

Critical weaknesses
  1. The provided method is neither fully supported nor new. The power sampling form can be consistently found in the long-tailed classification papers, and the proposed two-stage process is not new as it is even provided as a function in some libraries (see RandCropByLabelClassesd in MONAI [1] ). The only new claim is the q: 0->1->0 design, but there can be infinite number of q formulations that could be compared against, and there are insufficient evidence that such curriculum is the optimal, nor discussion on under what conditions such curriculum works well.

  2. The evaluation is insufficient and shows low performance. The best performance differs across different curriculum settings on each datasets, and the difference between them is also small, possibly in the statistical error bound. There are only few sampling-based baselines, but loss-reweighting baseline (such as focal loss) are much popularly used and therefore should not be out of the scope at the evaluation.

[1] https://monai.readthedocs.io/en/latest/_gen/monai.transforms.croppad.dictionary.html

Questions for authors

How can you prove that q: 0->1->0 design is optimal? Also, can you show the error bars of the experiments with additional seed runs and support the novelty of your method?

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

Reci-Q addresses imbalanced 3D medical image segmentation by selecting an anchor class according to voxel prevalence and cropping around an anchor voxel. A reciprocal schedule, q: 0 → 1 → 0, moves from class-balanced sampling to natural prevalence sampling and back. Experiments on five MSD tasks report improved overall and tail-class Dice over default nnU-Net sampling.

Critical strengths
  • Reci-Q controls class sampling through a single q while retaining nnU-Net’s network and soft Dice and cross-entropy losses, without auxiliary models or online difficulty estimation.

  • Reci-Q improves every designated tail class over nnU-Net across the five MSD tasks, increasing average tail-class Dice from 60.98 to 62.61.

Critical weaknesses
  • The implicit loss-weighting analysis uses the probability of selecting an anchor class as the weight of that class’s voxel-wise losses. A sampled patch can contain multiple classes, so a voxel’s effective weight depends on its inclusion in patches selected through different anchor classes. Figure 1 also changes the normalized background-to-foreground loss ratio at q=0, although all classes then have equal sampling probabilities. The weighted loss composition bars therefore do not follow the stated formulation and cannot substantiate the proposed reweighting mechanism.

  • The sampling formula pc∝ncqp_c \propto n_c^q already appears in Kang et al. [1], and MONAI’s class-based cropping already selects a class according to specified ratios and then a voxel from that class. Using per-volume voxel counts and introducing the reciprocal schedule are distinct changes, but presenting the separation of class selection and spatial localization as a new sampling mechanism overstates the methodological novelty.

  • The static comparisons include only q=0 and q=1, leaving the benefit over a fixed intermediate q unresolved. Both monotonic directions and different phase allocations are evaluated, but Reci-Q’s overall Dice is only 0.35 points above q: 0 → 1, with no reported variability. The evidence for a reliable advantage from the reciprocal trajectory itself therefore remains limited.

  • The PGPS comparison reports non-convergence on Task6, Task7, and Task8, yet the discussion describes Task6 as stable. Task6 is binary, so the explanation based on unequal prevalence among foreground classes does not account for its reported failure. With PGPS failing to converge on three of the five tasks, the comparison does not establish Reci-Q’s advantage over a functioning alternative curriculum.

[1] Decoupling Representation and Classifier for Long-Tailed Recognition.

Questions for authors

Please see my "weaknesses" comments.

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

Related papers

Discussion (0)

Sign in to comment.