Related Experiment Video
Updated: Nov 4, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.1K
Uncertainty-aware temporal self-learning (UATS): Semi-supervised learning for segmentation of prostate zones and
Anneke Meyer1, Suhita Ghosh1, Daniel Schindele2
1Faculty of Computer Science and Research Campus STIMULATE, University of Magdeburg, Germany.
Artificial Intelligence in Medicine
|May 22, 2021
Summary
This study introduces uncertainty-aware temporal self-learning (UATS) for precise prostate segmentation, improving accuracy for transition zone (TZ), peripheral zone (PZ), and other structures. The novel semi-supervised learning method achieves human-level performance, even with limited labeled data.
Area of Science:
- Medical image analysis
- Artificial intelligence in radiology
- Computational anatomy
Background:
- Convolutional neural networks (CNNs) show promise for prostate segmentation.
- Fine-grained segmentation of prostate zones (transition zone, peripheral zone, distal prostatic urethra, anterior fibromuscular stroma) remains challenging.
- Limited labeled data hinders supervised learning for detailed prostate segmentation.
Purpose of the Study:
- To develop a semi-supervised learning (SSL) method for accurate, fine-grained prostate segmentation.
- To overcome limitations of manual ground truth labeling in medical imaging.
- To improve segmentation performance using readily available unlabeled data.
Main Methods:
- Proposed uncertainty-aware temporal self-learning (UATS), combining temporal ensembling and uncertainty-guided self-learning.
- Applied SSL to leverage unlabeled images alongside limited labeled data.
- Evaluated UATS on prostate segmentation and generalized to hippocampus and skin lesion segmentation tasks.
Main Results:
- UATS significantly outperformed supervised baselines in segmenting prostate zones.
- Achieved Dice coefficients (DCs) up to 78.9% (TZ), 87.3% (PZ), 75.3% (DPU), and 50.6% (AFS).
- Demonstrated robustness against noise, generalization across varying labeled data ratios, and comparable performance to human inter-rater variability.
Conclusions:
- UATS effectively enhances fine-grained prostate segmentation accuracy using semi-supervised learning.
- The method shows superior performance, especially with minimal labeled data, addressing a key challenge in medical image analysis.
- UATS demonstrates broad applicability and robustness, offering a promising approach for complex segmentation tasks.
