Robust Semi-Supervised 3D Medical Image Segmentation With Diverse Joint-Task Learning and Decoupled Inter-Student...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised segmentation is crucial for 3D medical imaging.
• Existing teacher-student models struggle with limited data diversity, impacting consistency constraints.
• Model weight synchronization can further reduce effectiveness in low-data scenarios.

Purpose of the Study:

• To improve the reliability of consistency constraints in semi-supervised 3D medical image segmentation.
• To introduce a novel multi-student architecture that leverages decoupled knowledge.
• To enhance segmentation performance despite data scarcity.

Main Methods:

• Employed masked image modeling to construct diverse joint-tasks, enhancing consistency constraint reliability.
• Developed a novel architecture with one teacher and multiple students, utilizing varied masked inputs.
• Implemented concurrent segmentation and reconstruction tasks for complementary feature learning.
• Introduced inter-student learning among multiple students with shared encoding but distinct decoding.

Main Results:

• The proposed approach outperformed six mainstream semi-supervised methods across four medical datasets.
• Achieved superior Dice and Jaccard index improvements compared to the most competitive method on an in-house dataset.
• Demonstrated enhanced segmentation accuracy and reliability through diverse joint-task learning.

Conclusions:

• The novel multi-student architecture effectively addresses limitations of traditional dual-model approaches in semi-supervised medical image segmentation.
• Diverse joint-task learning via masked image modeling significantly enhances segmentation performance.
• The method shows strong potential for improving 3D medical image analysis in data-scarce environments.