Machine learning automatically detects COVID-19 using chest CTs in a large multicenter cohort.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Infectious Disease Diagnostics

Background:

• Chest computed tomography (CT) is crucial for diagnosing COVID-19.
• Distinguishing COVID-19 from other pulmonary conditions like pneumonia and interstitial lung disease (ILD) can be challenging.
• Machine learning (ML) and deep learning (DL) offer potential for automated analysis of CT scans.

Purpose of the Study:

• To evaluate ML and interpretable models for COVID-19 detection using chest CT.
• To differentiate COVID-19 from other pneumonias, ILD, and normal CT findings.
• To assess the performance of a DL-based classifier against interpretable ML models.

Main Methods:

• Retrospective multi-institutional study with 2446 chest CTs (1161 COVID-19 positive).
• Trained logistic regression and random forest models on interpretable features.
• Developed a DL classifier using 3D features from CT attenuation and opacity distribution.
• Utilized unsupervised hierarchical clustering to identify key features.

Main Results:

• Key COVID-19 features: percentage of airspace opacity, peripheral and basal predominant opacities.
• DL classifier achieved AUC=0.93 (sensitivity 90%, specificity 83%).
• Metrics-based classifier achieved AUC=0.83 (sensitivity 74%, specificity 79%).
• High accuracy for non-COVID-19 cases: ILD (91%), no pathologies (94%), other pneumonias (64%).

Conclusions:

• Quantitative imaging features from chest CT can accurately discriminate COVID-19.
• The DL-based method balances interpretability and classification performance.
• This approach may aid in COVID-19 diagnosis, especially in resource-limited settings.