Semi-supervised Strong-Teacher Consistency Learning for few-shot cardiac MRI image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiovascular disease is a major global health concern.
• Accurate segmentation of cardiac structures in MRI is vital for diagnosis.
• Supervised learning for segmentation requires extensive labeled data, which is scarce for cardiac MRI.

Purpose of the Study:

• To develop a semi-supervised learning model for cardiac MRI segmentation.
• To address the challenge of limited labeled data in cardiac MRI analysis.
• To improve the accuracy and efficiency of automated cardiac MRI segmentation.

Main Methods:

• Introduced a novel semi-supervised Strong-Teacher Consistency Network for few-shot multi-class cardiac MRI segmentation.
• Employed a student-teacher architecture with a multi-teacher structure to capture diverse perspectives.
• Utilized a hybrid loss function combining consistency and supervised losses, along with feature-space virtual adversarial training.

Main Results:

• The proposed model outperformed nine state-of-the-art semi-supervised methods on MM-WHS and ACDC datasets.
• Achieved 90.14% accuracy on MM-WHS with 25% labeling and 78.45% accuracy on ACDC with 1% labeling.
• Demonstrated superior performance compared to fully-supervised and single-teacher approaches, especially with limited annotations.

Conclusions:

• The Strong-Teacher Consistency Network effectively leverages unlabeled data for robust cardiac MRI segmentation.
• This approach significantly improves segmentation accuracy in few-shot scenarios with limited annotated data.
• The model offers a promising solution for automated cardiac MRI analysis, aiding in cardiovascular disease diagnostics.