Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Striving for Simplicity: Simple Yet Effective Prior-Aware Pseudo-Labeling for Semi-Supervised Ultrasound Image Segmentation

Yaxiong Chen, Yujie Wang, Zixuan Zheng, Jingliang Hu, Yilei Shi, Shengwu Xiong, Xiao Xiang Zhu, Lichao Mou

TL;DR

This work tackles the challenge of semi-supervised ultrasound image segmentation with limited labels by introducing a simple yet effective framework that leverages a shape prior. The model uses an encoder $\mathcal{E}$ and twin decoders $\mathcal{D}_l$ and $\mathcal{D}_p$, with a deep shape regularizer (DSR) learned from a pre-trained GAN discriminator to restrict anatomically implausible outputs without stifling ground-truth deviations. The training objective combines a supervised term $\mathcal{L}_s$ and an unsupervised term $\mathcal{L}_u$ (with a shape-regularization component $\mathcal{L}_{\mathrm{dsr}}$), yielding $\mathcal{L}=\mathcal{L}_s+\gamma\mathcal{L}_u$. Experiments on two public ultrasound benchmarks (TN3K and BUSI) show state-of-the-art performance under different label partitions, with notable gains when labels are scarce, and the approach provides a strong, lightweight baseline for future semi-supervised medical image segmentation tasks; code is publicly available.

Abstract

Medical ultrasound imaging is ubiquitous, but manual analysis struggles to keep pace. Automated segmentation can help but requires large labeled datasets, which are scarce. Semi-supervised learning leveraging both unlabeled and limited labeled data is a promising approach. State-of-the-art methods use consistency regularization or pseudo-labeling but grow increasingly complex. Without sufficient labels, these models often latch onto artifacts or allow anatomically implausible segmentations. In this paper, we present a simple yet effective pseudo-labeling method with an adversarially learned shape prior to regularize segmentations. Specifically, we devise an encoder-twin-decoder network where the shape prior acts as an implicit shape model, penalizing anatomically implausible but not ground-truth-deviating predictions. Without bells and whistles, our simple approach achieves state-of-the-art performance on two benchmarks under different partition protocols. We provide a strong baseline for future semi-supervised medical image segmentation. Code is available at https://github.com/WUTCM-Lab/Shape-Prior-Semi-Seg.

Striving for Simplicity: Simple Yet Effective Prior-Aware Pseudo-Labeling for Semi-Supervised Ultrasound Image Segmentation

TL;DR

This work tackles the challenge of semi-supervised ultrasound image segmentation with limited labels by introducing a simple yet effective framework that leverages a shape prior. The model uses an encoder and twin decoders and , with a deep shape regularizer (DSR) learned from a pre-trained GAN discriminator to restrict anatomically implausible outputs without stifling ground-truth deviations. The training objective combines a supervised term and an unsupervised term (with a shape-regularization component ), yielding . Experiments on two public ultrasound benchmarks (TN3K and BUSI) show state-of-the-art performance under different label partitions, with notable gains when labels are scarce, and the approach provides a strong, lightweight baseline for future semi-supervised medical image segmentation tasks; code is publicly available.

Abstract

Medical ultrasound imaging is ubiquitous, but manual analysis struggles to keep pace. Automated segmentation can help but requires large labeled datasets, which are scarce. Semi-supervised learning leveraging both unlabeled and limited labeled data is a promising approach. State-of-the-art methods use consistency regularization or pseudo-labeling but grow increasingly complex. Without sufficient labels, these models often latch onto artifacts or allow anatomically implausible segmentations. In this paper, we present a simple yet effective pseudo-labeling method with an adversarially learned shape prior to regularize segmentations. Specifically, we devise an encoder-twin-decoder network where the shape prior acts as an implicit shape model, penalizing anatomically implausible but not ground-truth-deviating predictions. Without bells and whistles, our simple approach achieves state-of-the-art performance on two benchmarks under different partition protocols. We provide a strong baseline for future semi-supervised medical image segmentation. Code is available at https://github.com/WUTCM-Lab/Shape-Prior-Semi-Seg.

Paper Structure

This paper contains 17 sections, 5 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Semi-Supervised Image Segmentation Network
  4. Shape Prior Modeling and Shape Regularization
  5. Shape Prior Modeling
  6. DSR
  7. Training Strategy and Loss
  8. Experiments
  9. Datasets and Evaluation Metrics
  10. Implementation Details
  11. DSR Training
  12. Segmentation Network Training
  13. Comparison with State-of-the-Art Methods
  14. Ablation Study
  15. Conclusion
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Examples of unrealistic and unnatural segmentations produced by a UNet.
  • Figure 2: Pipeline of the proposed semi-supervised method for ultrasound image segmentation.
  • Figure 3: Qualitative comparison of segmentation results produced by our method and other competitors on the TN3K and BUSI test sets. The top two rows show examples from TN3K, while the bottom two rows show cases from BUSI.
  • Figure 4: Qualitative assessment of our proposed DSR module on the TN3K and BUSI test sets. The top row shows an example from TN3K, while the bottom case is from BUSI.