Active contour regularized semi-supervised learning for COVID-19 CT infection segmentation with limited annotations.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Accurate COVID-19 infection segmentation on chest CT is crucial for quantitative analysis.
• Developing automated segmentation tools is urgent due to limited labeled data and short timeframes.
• Pseudo-labeling is a promising semi-supervised approach but can yield inaccurate labels.

Purpose of the Study:

• To propose an active contour regularized semi-supervised learning framework for automatic COVID-19 infection segmentation.
• To leverage limited labeled images and unlabeled data for improved segmentation performance.
• To address the challenge of inaccurate pseudo-labels in existing methods.

Main Methods:

• Developed a semi-supervised learning framework incorporating active contour regularization via the region-scalable fitting (RSF) model.
• Embedded the RSF model into the network's loss function to regularize and refine pseudo-labels.
• Implemented a splitting method for separate optimization of RSF and segmentation loss terms using iterative convolution-thresholding and stochastic gradient descent.

Main Results:

• The proposed method significantly outperforms state-of-the-art techniques on both small and large datasets.
• Achieved up to 5% improvement in Dice Similarity Coefficient and Normalized Surface Dice.
• Demonstrated up to 10% reduction in Relative Absolute Volume Difference and 8 mm in 95% Hausdorff Distance.

Conclusions:

• The active contour regularized semi-supervised framework effectively segments COVID-19 infections with limited labeled data.
• Identified novel infection patterns in dorsal subpleural lung and posterior basal segments, advancing COVID-19 understanding.
• The method offers a robust solution for automated infection segmentation in medical imaging.