A General Global and Local Pre-Training Framework for 3D Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate medical image segmentation is vital for surgical robot precision.
• Limited annotated medical data hinders the development of robust segmentation models.
• Existing self-supervised methods struggle to leverage both global and local image features.

Purpose of the Study:

• To develop a novel self-supervised pre-training framework for volumetric medical image segmentation.
• To address the challenge of limited annotations in medical imaging.
• To improve the accuracy of target segmentation for surgical applications.

Main Methods:

• Proposed a pre-training framework utilizing 3D anatomical structures and task-specific cues.
• Introduced multiple sub-tasks to learn intrinsic patterns of volumetric medical image structures.
• Designed a multi-level background cube contrastive learning strategy to enhance feature representation.

Main Results:

• Achieved significant improvements over existing self-supervised learning techniques under limited annotation settings.
• Demonstrated performance within 6% of baseline using only five labeled CT volumes.
• Extensive evaluations conducted on two publicly available datasets.

Conclusions:

• The proposed self-supervised method effectively segments volumetric medical images with limited annotations.
• The approach successfully integrates global anatomical information and local feature differences.
• This work paves the way for more robust and data-efficient medical image segmentation models.