COSST: Multi-Organ Segmentation With Partially Labeled Datasets Using Comprehensive Supervisions and Self-Training

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep learning excels at multi-organ segmentation but requires extensive, fully annotated datasets.
• Medical datasets are often limited in size and only partially labeled, posing a challenge for unified model training.

Purpose of the Study:

• To investigate methods for learning unified models from partially labeled medical image datasets.
• To leverage the synergistic potential of multiple partially labeled datasets for improved segmentation.

Main Methods:

• Proposed a novel two-stage framework, COSST (Comprehensive Supervision Signals with Self-Training).
• Integrated two ground truth-based signals and one pseudo-label signal using self-training.
• Assessed pseudo-label reliability via outlier detection in latent space to mitigate performance degradation.

Main Results:

• COSST achieved significant improvements over baseline methods (individual networks per dataset).
• Demonstrated consistent superior performance compared to state-of-the-art partial-label segmentation methods.
• Validated on 12 CT datasets across public and private partial-label segmentation tasks.

Conclusions:

• COSST effectively and efficiently integrates comprehensive supervision signals with self-training for partial-label segmentation.
• The framework successfully addresses the challenge of learning from limited and partially annotated medical image data.
• COSST offers a robust solution for improving multi-organ segmentation accuracy in real-world scenarios.