Introduction
Chest CT imaging is a cornerstone for diagnosing and monitoring lung infections. However, radiological assessments are traditionally qualitative, relying on a radiologist's ability to identify features like ground-glass opacities (GGO) or consolidations. There is a significant lack of computational tools to quantify the exact spatial distribution of these infections across populations or to track their precise evolution over time.
Understanding these patterns at scale is essential for distinguishing novel diseases like COVID-19 from standard community-acquired pneumonia (CAP), tracking disease progression, and identifying imaging biomarkers for patient outcomes. This paper introduces an automated deep learning pipeline designed to segment and register thousands of 3D CT scans, enabling a comprehensive population-level analysis.
Methods
The researchers developed a three-step automated pipeline to generate population-level infection distribution maps.
- Segmentation (VB-Net): A modified 3D convolutional neural network, VB-Net, was used to segment lung fields and infection regions. To optimize training, a Human-Involved-Model-Iterations (HIMI) strategy was employed, where radiologists corrected model pre-segmentations, drastically reducing annotation time from 211 minutes to 4.7 minutes per case.
- Registration: Segmented lung masks were warped onto a common "template" CT image using affine and deformable registration based on lung boundaries, allowing for direct spatial comparison across different patients.
- Distribution Mapping: Voxel-level probability maps were calculated by overlaying warped masks. Infections were further categorized into GGO and consolidation based on Hounsfield Unit (HU) thresholds (GGO < -300 HU; consolidation ≥ -300 HU).
Key Results and Insights
The study utilized a massive dataset including nearly 3,000 COVID-19 images and over 1,300 CAP images. The VB-Net achieved a high segmentation accuracy with a Dice coefficient of 91.6% ± 10.0%, comparable to inter-rater variability between experts.
Analysis revealed that COVID-19 has a distinct "spatial fingerprint." Unlike CAP, COVID-19 is more widely distributed and favors peripheral, subpleural regions, particularly in the right lower lobe. Statistically significant differences in distribution were found in 48.0% of the lung volume. Furthermore, COVID-19 is dominated by GGOs; when consolidations occur, they are located even further toward the lung periphery than GGOs.
Longitudinal analysis of severe and critical patients provided two major findings. First, the spatial distribution of lesions at disease onset is nearly identical between those who eventually require ICU admission and those who do not, explaining the difficulty of early severity prediction. Second, critical patients exhibit a "double-dip" recovery pattern: an initial progression, a period of absorption (days 6-13), a secondary enlargement phase (days 14-17), and finally, slow absorption.
Conclusion
This automated pipeline successfully quantifies the spatial and temporal characteristics of lung infections at scale. While the current registration method relies on lung surfaces and could be improved by incorporating internal anatomical landmarks, the resulting distribution maps provide a powerful tool for disease characterization. These spatial and HU-based features offer a foundation for developing future diagnostic classifiers to assist in clinical decision-making.
Summary updated