Uncertainty-aware multi-view co-training for semi-supervised medical image segmentation and domain adaptation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Deep learning for medical image segmentation requires extensive annotated data, which is costly and time-consuming to obtain.
• Unlabeled medical data is abundant and can be leveraged to improve segmentation performance.
• Semi-supervised learning and unsupervised domain adaptation are crucial for utilizing unlabeled data.

Purpose of the Study:

• To propose a unified framework, uncertainty-aware multi-view co-training (UMCT), for volumetric medical image segmentation.
• To efficiently utilize unlabeled data for enhanced performance in both semi-supervised and unsupervised domain adaptation settings.
• To address the challenge of limited annotated data in medical image analysis.

Main Methods:

• Developed a unified framework (UMCT) for semi-supervised learning and unsupervised domain adaptation in medical image segmentation.
• Employed multi-view representation by rotating and permuting 3D volumes.
• Implemented co-training with uncertainty estimation for accurate labeling of unlabeled data.

Main Results:

• Achieved state-of-the-art performance on semi-supervised medical image segmentation tasks using the NIH pancreas and a multi-organ dataset.
• Demonstrated effectiveness in unsupervised domain adaptation by adapting models to pathological organs.
• Validated the framework's ability to handle scenarios with inaccessible labeled source data.

Conclusions:

• The proposed UMCT framework significantly enhances medical image segmentation by effectively leveraging unlabeled data.
• UMCT shows strong potential for real-world applications, particularly in unsupervised domain adaptation scenarios.
• The uncertainty-aware multi-view co-training approach offers a robust solution for data-scarce medical imaging challenges.