COVID-19 Infection Segmentation and Severity Assessment Using a Self-Supervised Learning Approach.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Automated segmentation and severity assessment of COVID-19 lesions are crucial for patient care.
• Deep learning models, while effective, require substantial annotated medical image data, which is often scarce.
• Training deep convolutional neural network (CNN) models is challenging due to the large volume of medical image samples.

Purpose of the Study:

• To develop a novel self-supervised deep learning method for automated COVID-19 lesion segmentation and severity assessment.
• To reduce the dependency on large amounts of annotated training samples for COVID-19 medical image analysis.
• To improve the efficiency and accuracy of COVID-19 diagnosis and treatment planning.

Main Methods:

• A self-supervised deep learning approach utilizing an encoder-decoder model.
• Unlabeled data is used for pre-training to learn rotation-dependent and rotation-invariant features.
• A small subset of labeled data is employed for fine-tuning the model for severity classification and lesion segmentation.

Main Results:

• The proposed method achieved superior COVID-19 severity classification accuracy compared to other unsupervised methods (7.16% higher).
• Segmentation performance, measured by Dice coefficient, showed significant improvements over U-Net, especially with limited labeled data (up to 16.88% higher).
• The method was validated on public and self-built COVID-19 CT datasets.

Conclusions:

• The proposed self-supervised deep learning method offers enhanced classification and segmentation performance for COVID-19 analysis.
• This approach effectively addresses the challenge of limited labeled data in medical image analysis.
• The method demonstrates superior performance compared to existing techniques, particularly in resource-constrained scenarios.