COVID TV-UNet: Segmenting COVID-19 Chest CT Images Using Connectivity Imposed U-Net

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• The COVID-19 pandemic necessitated rapid diagnostic tools, with Computed Tomography (CT) scans emerging as a viable alternative to RT-PCR testing.
• Accurate segmentation of infected lung regions in CT images is crucial for diagnosis and patient management.

Approach:

• A novel segmentation framework, TV-UNet, based on the U-Net architecture was developed for pixel-level detection of COVID-19 related ground glass opacities.
• A 2D-anisotropic total-variation regularization term was incorporated into the loss function to encourage connectivity in segmented regions, reflecting the nature of viral spread.

Key Points:

• The TV-UNet model demonstrated a 2% improvement in overall segmentation performance compared to the standard U-Net model on a dataset of approximately 900 CT images.
• Quantitative assessments showed high performance, with a mean Intersection over Union (mIoU) rate exceeding 99% and a Dice score around 86% for identifying COVID-19 affected areas.
• The model effectively identifies COVID-19 associated regions within the lungs, showcasing its potential for clinical application.

Conclusions:

• The TV-UNet model offers a robust and accurate method for segmenting COVID-19 infected regions in chest CT images.
• The integration of a connectivity-promoting regularization term significantly enhances segmentation performance, outperforming traditional U-Net architectures.
• This AI-driven approach provides a valuable tool for the rapid and precise detection of COVID-19 in radiological imaging.