Related Experiment Video
Updated: Aug 5, 2025
08:05
Lung CT Segmentation to Identify Consolidations and Ground Glass Areas for Quantitative Assesment of SARS-CoV Pneumonia
Published on: December 19, 2020
14.2K
Deep-Learning-Based Whole-Lung and Lung-Lesion Quantification Despite Inconsistent Ground Truth: Application to
Syed M S Reza1, Winston T Chu1, Fatemeh Homayounieh1
1Center for Infectious Disease Imaging, Radiology and Imaging Sciences, Clinical Center, National Institutes of Health, Bethesda, Maryland.
Academic Radiology
|March 25, 2023
Summary
We developed an automated deep-learning system for analyzing lung CT scans in nonhuman primate models of COVID-19. This novel approach accurately quantifies lung lesions, improving preclinical research for infectious diseases.
Area of Science:
- Preclinical research
- Infectious disease modeling
- Medical imaging analysis
Background:
- Animal models are crucial for studying infectious diseases like COVID-19, enabling controlled experiments and longitudinal measurements.
- Computerized tomography (CT) imaging is vital for disease characterization, but automated analysis using deep learning has faced limitations due to manual segmentation inconsistencies.
- Applying deep learning to nonhuman primate (NHP) models of SARS-CoV-2 presents unique challenges in translating human-derived tools.
Purpose of the Study:
- To develop and evaluate a deep-learning-based quantification method for whole lungs and lung lesions on CT scans in NHP models of SARS-CoV-2.
- To address the challenge of inconsistent ground truth data in automated segmentation through a novel multi-model ensemble technique.
- To improve the accuracy and standardization of disease detection and quantification in preclinical infectious disease research.
Main Methods:
- Utilized a deep-learning approach, specifically a convolutional neural network (CNN) with a feature pyramid network (FPN), for automated segmentation of lungs and lesions.
- Implemented a multi-model ensemble technique, training multiple CNNs on different data subsets to mitigate ground truth inconsistencies.
- Employed an FPN to handle variations in object size, improving prediction accuracy across different scales.
Main Results:
- Achieved high Dice coefficients: 99.4% for whole-lung segmentation and 60.2% for lung-lesion segmentation.
- The proposed multi-model FPN significantly outperformed established methods like U-Net, V-Net, and Inception for lung-lesion segmentation.
- Demonstrated the utility of the segmentation outputs for longitudinal quantification of lung disease in SARS-CoV-2-exposed NHPs.
Conclusions:
- Deep learning offers a standardized and automated method for detecting and quantifying lung disease in preclinical NHP models.
- The developed multi-model FPN technique effectively addresses segmentation inconsistencies, enhancing the reliability of automated analysis.
- Future deep-learning applications in preclinical research should be specifically tailored to address unique needs for impact, automation, and dynamic quantification.
