Dual U-Net Approach for Automated Lung and COVID-19 Lesion Segmentation in CT

Overview

This study presents LungQuant, a fully automated deep learning system using a dual U-Net architecture to segment lungs and COVID-19 lesions on CT scans. The system was trained and evaluated across multiple publicly available datasets with varying annotation standards, demonstrating its generalization ability and quantifying lung involvement via CT severity scores.

Background

COVID-19 pneumonia manifests with heterogeneous lung abnormalities on CT, including ground-glass opacities and consolidations, which are challenging to segment manually due to their variable appearance. Deep learning methods, particularly U-Net architectures, have shown promise in automating this segmentation task but require sufficient annotated data. Variability in dataset acquisition protocols and annotation standards complicates model training and evaluation. LungQuant addresses these challenges by leveraging multiple datasets and a dual-stage U-Net pipeline to segment lungs and lesions, providing quantitative assessment of lung involvement.

Data Highlights

Key Findings

• The LungQuant system uses a cascade of two U-Nets: the first segments lungs, the second segments COVID-19 lesions within the lung region.
• Training on heterogeneous datasets with different annotation standards improves the model's generalization ability across varied CT data.
• The system outputs lung and lesion masks, calculates the percentage of lung volume affected, and assigns a CT severity score (CT-SS) based on lesion burden.
• Data harmonization is limited by missing acquisition metadata, but training on large, diverse datasets partially overcomes this challenge.
• Conversion from DICOM to NifTI format and slice subsampling (e.g., in MosMed) can reduce resolution and impact segmentation quality.
• Residual connections and instance normalization in the 6-level deep U-Net architecture enhance segmentation performance.

Clinical Implications

The LungQuant system offers a reproducible, automated tool for quantifying lung involvement in COVID-19 pneumonia from CT scans, facilitating objective assessment of disease severity. Its ability to generalize across datasets with different annotation styles supports broader clinical applicability. This approach may aid radiologists in monitoring disease progression and guiding treatment decisions without requiring extensive manual segmentation.

Conclusion

The dual U-Net LungQuant pipeline effectively segments lungs and COVID-19 lesions across heterogeneous datasets, providing reliable quantification of lung involvement. This work underscores the importance of diverse training data and standardized evaluation for developing robust automated imaging tools in COVID-19.

References

1. Morozov et al. 2020 -- MosMed Dataset Description
2. Ma et al. 2020 -- Deep Learning for Lung and COVID-19 Lesion Segmentation
3. Lessmann et al. 2020 -- U-Net Model for COVID-19 Lesion Segmentation
4. Fang et al. 2020 -- Automated Lung and Lesion Segmentation Methods
5. Hofmanninger et al. 2020 -- Lung Segmentation Method
6. Ronneberger et al. 2015 -- U-Net: Convolutional Networks for Biomedical Image Segmentation