Learning Geometric Information Propagation for Semi-supervised 3D Medical Image Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning (SSL) is crucial for medical image segmentation, reducing the need for large labeled datasets.
• Existing methods often overlook global geometric priors in medical images, limiting segmentation performance.
• Distribution discrepancies between labeled and unlabeled data in SSL hinder model generalization.

Purpose of the Study:

• To propose a novel semi-supervised medical image segmentation method (GGIP) that leverages global geometric priors and addresses data distribution discrepancies.
• To enhance the utilization of global geometric information and improve model generalization in medical image segmentation.

Main Methods:

• Developed a geometric moment attention mechanism for extracting rich global geometric features.
• Implemented a global geometric perturbation consistency strategy to account for anatomical variations.
• Designed a global information propagation Mamba module to mitigate distribution discrepancies between labeled and unlabeled data.

Main Results:

• The proposed GGIP method achieved state-of-the-art (SOTA) performance on multiple medical image segmentation datasets.
• Demonstrated improved accuracy in segmenting pancreas, left atrium, and brain tumors.
• Successfully reduced distribution discrepancies and enhanced generalization capabilities.

Conclusions:

• The GGIP method effectively incorporates global geometric attention and information propagation for superior semi-supervised medical image segmentation.
• This approach offers a promising direction for improving medical image analysis by addressing limitations in existing SSL techniques.