BASIC: Semi-supervised Multi-organ Segmentation with Balanced Subclass Regularization and Semantic-conflict Penalty

Zhenghao Feng, Lu Wen, Yuanyuan Xu, Binyu Yan, Xi Wu, Jiliu Zhou, Yan Wang

Abstract

Semi-supervised learning (SSL) has shown notable potential in relieving the heavy demand of dense prediction tasks on large-scale well-annotated datasets, especially for the challenging multi-organ segmentation (MoS). However, the prevailing class-imbalance problem in MoS caused by the substantial variations in organ size exacerbates the learning difficulty of the SSL network. To address this issue, in this paper, we propose an innovative semi-supervised network with BAlanced Subclass regularIzation and semantic-Conflict penalty mechanism (BASIC) to effectively learn the unbiased knowledge for semi-supervised MoS. Concretely, we construct a novel auxiliary subclass segmentation (SCS) task based on priorly generated balanced subclasses, thus deeply excavating the unbiased information for the main MoS task with the fashion of multi-task learning. Additionally, based on a mean teacher framework, we elaborately design a balanced subclass regularization to utilize the teacher predictions of SCS task to supervise the student predictions of MoS task, thus effectively transferring unbiased knowledge to the MoS subnetwork and alleviating the influence of the class-imbalance problem. Considering the similar semantic information inside the subclasses and their corresponding original classes (i.e., parent classes), we devise a semantic-conflict penalty mechanism to give heavier punishments to the conflicting SCS predictions with wrong parent classes and provide a more accurate constraint to the MoS predictions. Extensive experiments conducted on two publicly available datasets, i.e., the WORD dataset and the MICCAI FLARE 2022 dataset, have verified the superior performance of our proposed BASIC compared to other state-of-the-art methods.

Introduction

Multi-organ segmentation (MoS) is a cornerstone of computer-assisted diagnosis, but it faces two major hurdles: the high cost of manual labeling and the inherent class-imbalance problem. In a typical abdominal CT scan, large organs like the liver dominate the volume, while small structures like the adrenal glands are tiny. Standard AI models tend to "over-focus" on large organs to minimize global error, often misidentifying smaller ones. This bias is exacerbated in semi-supervised learning (SSL) settings, where the model must generate pseudo-labels for unlabeled data.

To address this, we introduce BASIC (BAlanced Subclass regularIzation and semantic-Conflict penalty). BASIC ensures small organs aren't overshadowed by their larger neighbors by breaking organs into balanced subclasses and applying anatomical constraints.

Methodology

The BASIC framework is built on the Mean Teacher architecture, where a "student" model learns from both labeled data and a "teacher" model's predictions on unlabeled data. The framework introduces three key innovations:

1. Balanced Subclass Partition (BSP)

Instead of treating the liver as one giant class and the adrenal gland as one tiny class, BASIC uses a balanced clustering algorithm to divide large organs into multiple subclasses. This ensures that every subclass has roughly the same number of pixels, creating a balanced training field for the encoder.

2. Balanced Subclass Regularization (BSR)

The model performs two tasks simultaneously: the main MoS task and an auxiliary Subclass Segmentation (SCS) task. The teacher's SCS predictions (the "balanced" view) are used to supervise the student's MoS predictions. This transfers unbiased knowledge from the balanced subclasses back to the main organ-level task.

3. Semantic-conflict Penalty (SCP)

Subclasses are "children" of "parent" organs. If the model predicts a subclass that belongs to the wrong parent organ (e.g., a liver subclass appearing inside the spleen), the system applies a heavy penalty. This maintains the hierarchical logic of human anatomy.

Overview of the BASIC framework

Experiments and Results

BASIC was evaluated on two major benchmarks: the WORD Dataset (16 organs) and the MICCAI FLARE 2022 Dataset (13 organs). We simulated extreme low-data scenarios, testing the model with as few as 5 labeled volumes.

Quantitative Performance

BASIC consistently outperformed state-of-the-art (SOTA) methods like Mean Teacher (MT), UAMT, and ICT. On the WORD dataset with only 5 labeled samples, BASIC achieved an Average Dice score of 72.79%, beating the next best method (EVIL) by over 4%.

Quantitative performance on the WORD dataset

The most dramatic improvements were seen in small organs. At $n=5$ labeled samples, BASIC improved the Dice score for the Spleen by 8.00% and the Stomach by 18.59% compared to the second-best performing model. These results prove that granularity is a powerful cure for class bias in medical imaging.

Conclusion

The BASIC framework demonstrates that artificially breaking down large classes into smaller, balanced ones forces encoders to learn more nuanced features. By combining this with anatomical hierarchy constraints, BASIC significantly improves the reliability of semi-supervised multi-organ segmentation, particularly for small, clinically significant structures.

Summary updated

Abstract

Semi-supervised learning (SSL) has shown notable potential in relieving the heavy demand of dense prediction tasks on large-scale well-annotated datasets, especially for the challenging multi-organ segmentation (MoS). However, the prevailing class-imbalance problem in MoS caused by the substantial variations in organ size exacerbates the learning difficulty of the SSL network. To address this issue, in this paper, we propose an innovative semi-supervised network with BAlanced Subclass regularIzation and semantic-Conflict penalty mechanism (BASIC) to effectively learn the unbiased knowledge for semi-supervised MoS. Concretely, we construct a novel auxiliary subclass segmentation (SCS) task based on priorly generated balanced subclasses, thus deeply excavating the unbiased information for the main MoS task with the fashion of multi-task learning. Additionally, based on a mean teacher framework, we elaborately design a balanced subclass regularization to utilize the teacher predictions of SCS task to supervise the student predictions of MoS task, thus effectively transferring unbiased knowledge to the MoS subnetwork and alleviating the influence of the class-imbalance problem. Considering the similar semantic information inside the subclasses and their corresponding original classes (i.e., parent classes), we devise a semantic-conflict penalty mechanism to give heavier punishments to the conflicting SCS predictions with wrong parent classes and provide a more accurate constraint to the MoS predictions. Extensive experiments conducted on two publicly available datasets, i.e., the WORD dataset and the MICCAI FLARE 2022 dataset, have verified the superior performance of our proposed BASIC compared to other state-of-the-art methods.
BibTeX
@misc{feng2025basic,
  author = {Zhenghao Feng and Lu Wen and Yuanyuan Xu and Binyu Yan and Xi Wu and Jiliu Zhou and Yan Wang},
  title = {{BASIC: Semi-supervised Multi-organ Segmentation with Balanced Subclass Regularization and Semantic-conflict Penalty}},
  year = {2025},
  eprint = {2501.03580},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.03580},
}
Figure