Deep learning for head and neck semi-supervised semantic segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiation Oncology

Background:

• Radiation therapy (RT) is a primary treatment for head and neck (H&N) cancers.
• Accurate delineation of organs-at-risk (OARs) on computed tomography (CT) scans is critical for RT planning.
• Manual OAR delineation is time-consuming and susceptible to inter-observer variability and domain shift across institutions.

Purpose of the Study:

• To develop an automated method for precise OAR delineation in H&N cancer RT planning.
• To address challenges of domain shift and confirmation bias in semi-supervised learning for medical image segmentation.
• To leverage unlabeled CT data to improve model training and generalization.

Main Methods:

• Proposed a novel cross-domain orthogon-based-perspective consistency (CD-OPC) strategy within a two-branch collaborative training framework.
• Introduced a generative pretext task, cross-domain prediction (CDP), for learning inherent CT image properties.
• Utilized orthogon-based pseudo-labeling for knowledge transfer to enhance sub-network feature learning.

Main Results:

• The CD-OPC model demonstrated superior performance compared to other semi-supervised semantic segmentation algorithms across H&N datasets from nine institutions.
• The model was trained on multi-institutional data and validated on local datasets, showing advanced segmentation accuracy.
• Achieved more advanced performance than other semi-supervised semantic segmentation algorithms.

Conclusions:

• The CD-OPC method effectively mitigates domain shift and prevents network collapse in medical image segmentation.
• The approach enhances network perceptual abilities and generates more reliable predictions for OARs.
• Successfully addresses confirmation bias in semi-supervised learning for RT planning, improving prediction reliability.