Semi-supervised Medical Image Segmentation Using Heterogeneous Complementary Correction Network and Confidence...

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Semi-supervised learning is crucial for medical image segmentation, with the mean-teacher (MT) framework showing promise.
• Existing MT methods face limitations due to unreliable pseudo-labels generated by the teacher model.

Purpose of the Study:

• To propose an innovative semi-supervised method for medical image segmentation, addressing limitations of current approaches.
• To enhance the accuracy and robustness of medical image segmentation using a novel framework.

Main Methods:

• Developed a triple-branch framework integrating a heterogeneous complementary correction (HCC) network into the MT framework.
• Introduced confidence contrastive learning (CCL) with a novel sampling strategy for improved feature learning.
• Employed momentum style transfer (MST) and Cutout-style augmentation to bridge data distribution gaps and boost unsupervised performance.

Main Results:

• The proposed HC-CCL method demonstrated significant performance advantages over existing approaches.
• Achieved state-of-the-art performance across all metrics on three diverse medical image datasets (LA, NIH pancreas, Brats-2019).

Conclusions:

• HC-CCL effectively corrects prediction errors and enhances feature learning in semi-supervised medical image segmentation.
• The method offers a robust and high-performing solution for clinical diagnosis applications.