Complementary consistency semi-supervised learning for 3D left atrial image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Semi-supervised learning methods for medical image segmentation often struggle to fully utilize unlabeled data.
• Existing algorithms have limitations in extracting valuable information from unlabeled datasets, impacting segmentation performance.

Purpose of the Study:

• To introduce CC-Net, a novel network designed for semi-supervised left atrium image segmentation.
• To enhance the utilization of unlabeled data through complementary information and consistency training.

Main Methods:

• Developed CC-Net, featuring a complementary symmetrical structure with a main and two auxiliary models.
• Implemented model-level perturbation and enforced consistency between models to capture complementary information.
• Focused on ambiguous areas and low-uncertainty decision boundaries for improved segmentation.

Main Results:

• CC-Net demonstrated superior performance in semi-supervised segmentation compared to state-of-the-art algorithms on two public datasets.
• The network effectively utilized unlabeled data by focusing on complementary information and enforcing model consistency.
• Achieved best performance under specific proportions of annotated data.

Conclusions:

• CC-Net offers an efficient and effective approach for semi-supervised left atrium image segmentation.
• The complementary consistency training strategy significantly enhances the ability to extract information from unlabeled data.
• The proposed method provides a promising direction for improving medical image segmentation with limited labeled data.