Separated collaborative learning for semi-supervised prostate segmentation with multi-site heterogeneous unlabeled...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Prostate segmentation from MRI is crucial for cancer staging and treatment planning.
• Semi-supervised learning (SSL) is appealing for medical images due to limited labeled data.
• Data heterogeneity across multiple sites challenges existing SSL methods.

Purpose of the Study:

• To propose a novel framework for semi-supervised prostate segmentation using multi-site unlabeled MRI data.
• To address the challenges of data scarcity and heterogeneity in collaborative medical image analysis.
• To improve the generalizability and robustness of SSL in multi-center settings.

Main Methods:

• Introduced a Separated Collaborative Learning (SCL) framework based on a teacher-student model.
• Implemented local learning strategies including pseudo-labeling and cyclic propagated real-label learning.
• Incorporated external multi-site learning with mutual dependence and adversarial perturbation stability learning.

Main Results:

• The SCL framework effectively generalizes semi-supervised learning to multi-site unlabeled MRI data.
• Achieved significant performance improvements over existing semi-supervised segmentation methods.
• Demonstrated extensibility to multi-class cardiac MRI segmentation across multiple centers.

Conclusions:

• The proposed SCL framework offers a robust solution for semi-supervised medical image segmentation with multi-site unlabeled data.
• This approach effectively overcomes data heterogeneity challenges in collaborative learning scenarios.
• The method shows promise for improving prostate cancer staging and treatment planning through enhanced MRI segmentation.