Quality-aware semi-supervised learning for CMR segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning for medical image segmentation faces challenges due to limited annotated training data.
• Existing methods like data augmentation and semi-supervised learning (SSL) have limitations in effectiveness.
• Segmentation is often an intermediate step for downstream clinical tasks, requiring robust analysis.

Purpose of the Study:

• To propose a novel semi-supervised learning (SSL) scheme, semiQCSeg, that leverages quality control (QC) of downstream tasks for improved medical image segmentation.
• To enhance the training of segmentation networks by identifying high-quality outputs for further training, effectively creating quality-aware data augmentation.

Main Methods:

• Developed semiQCSeg, a variant of SSL for segmentation networks, using QC of downstream tasks to select high-quality segmentation outputs.
• Applied the method to aortic and short axis cardiac volume segmentation tasks using UK Biobank data.
• Evaluated performance using U-net and Fully Convolutional Network architectures, comparing against supervised and standard SSL strategies.

Main Results:

• SemiQCSeg demonstrated improved training of segmentation networks compared to supervised and standard SSL methods.
• The approach reduced the requirement for labeled data.
• Achieved superior performance in both Dice and clinical metrics for cardiac MRI segmentation tasks.

Conclusions:

• SemiQCSeg offers an efficient strategy for training medical image segmentation networks, particularly when labeled datasets are scarce.
• Leveraging downstream task QC provides a robust method for quality-aware data augmentation in semi-supervised settings.
• The proposed method enhances segmentation accuracy and reduces reliance on extensive manual annotation.