Uncertainty-Aware Semi-Supervised Method for Pectoral Muscle Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning methods use unlabeled data for improved model performance.
• Poor-quality predictions in regularization terms lead to noisy gradients and degraded performance.
• Existing methods filter low-confidence predictions, hindering learning from uncertain data regions.

Purpose of the Study:

• To develop an uncertainty-aware semi-supervised method for pectoral muscle segmentation in medical imaging.
• To improve the quality of target predictions and enhance the utilization of unlabeled data.
• To address the challenge of low-confidence predictions in cross-domain scenarios.

Main Methods:

• Proposed an uncertainty-aware semi-supervised method with a teacher-student dual model structure.
• Incorporated breast anatomical prior for pectoral muscle segmentation.
• Designed a low-confidence prediction refinement module using high-confidence predictions and learned anatomical prior.

Main Results:

• Achieved an average improvement of 1.76 in DICE index compared to the baseline method.
• Demonstrated an average reduction of 3.21 in IoU index and 5.48 in HD index.
• Showcased strong generalization performance across three data centers and outperformed other methods.

Conclusions:

• The proposed uncertainty-aware method effectively refines low-confidence predictions using anatomical priors.
• This approach improves pectoral muscle segmentation accuracy and robustness in semi-supervised learning.
• The method offers a significant advancement for medical image analysis, particularly in challenging segmentation tasks.