Related Experiment Video
Updated: Sep 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
14.3K
Unsupervised segmentation and quantification of COVID-19 lesions on computed Tomography scans using CycleGAN
Marc Connell1, Yi Xin2, Sarah E Gerard3
1Department of Anesthesiology and Critical Care, University of Pennsylvania, Philadelphia, PA, USA.
Methods (San Diego, Calif.)
|July 11, 2022
Summary
This study introduces a novel method for automated COVID-19 lung lesion segmentation using CycleGAN, eliminating the need for manual data labeling. The approach successfully identifies and quantifies pathological tissue in CT scans, offering a valuable tool for pandemic response and resource-limited settings.
Area of Science:
- Medical image analysis
- Artificial intelligence in radiology
- Computational pathology
Background:
- Lesion segmentation in medical imaging is crucial for disease identification.
- Manual data labeling is time-intensive and impractical in resource-limited settings or pandemics.
- Automated methods are needed for efficient COVID-19 lung tissue analysis.
Purpose of the Study:
- To develop a fully automated method for segmenting and quantifying pathological COVID-19 lung tissue.
- To eliminate the requirement for manually segmented training data in medical image analysis.
- To create a CycleGAN-based approach for converting COVID-19 CT scans to healthy equivalents.
Main Methods:
- A cycle-consistent generative adversarial network (CycleGAN) was trained to transform COVID-19 CT scans into generated healthy images.
- Pathological tissue maps were created by subtracting generated healthy images from original scans.
- Three-dimensional lesion segmentations were constructed from these pathological maps.
- Segmentation performance was evaluated using Dice scores against radiologist-reviewed ground truth.
Main Results:
- The CycleGAN generator effectively removed high Hounsfield unit (HU) voxels representing lesions, replacing them with lower HU values.
- Normal anatomical structures like vessels and airways were preserved without distortion.
- Generated healthy images showed increased gas content and decreased tissue density compared to COVID-19 images.
- Lesion segmentations achieved an average Dice score of 55.9, comparable to other weakly supervised methods.
Conclusions:
- The CycleGAN model successfully segmented pulmonary lesions in both mild and severe COVID-19 cases.
- This automated segmentation method offers comparable performance to existing models.
- The key innovation is the ability to segment lesions without manual segmentation, addressing critical needs in medical analysis.
