A fully automatic deep learning system for COVID-19 diagnostic and prognostic analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• Global spread of Coronavirus disease 2019 (COVID-19) has strained medical resources.
• Efficient diagnosis and risk stratification are crucial for patient management and resource optimization.
• Computed tomography (CT) is a key imaging modality for respiratory illness assessment.

Purpose of the Study:

• To develop and validate a fully automatic deep learning system for COVID-19 diagnosis using CT.
• To assess the system's ability to predict patient prognosis and stratify risk.
• To evaluate the system's utility in clinical settings for rapid screening and resource allocation.

Main Methods:

• Retrospective collection of 5372 CT images from patients across seven cities/provinces.
• A deep learning system was pre-trained on 4106 images to learn lung features.
• The system was trained and externally validated on 1266 patients (924 COVID-19, 342 other pneumonia).

Main Results:

• The deep learning system demonstrated strong performance in identifying COVID-19 from other pneumonia (AUCs 0.87-0.88) and viral pneumonia (AUC 0.86).
• The system successfully stratified patients into high- and low-risk groups with significant differences in hospital stay duration (p=0.013, p=0.014).
• The automated system focused on abnormal lung areas consistent with radiological findings without human intervention.

Conclusions:

• Deep learning offers a valuable tool for rapid COVID-19 screening via CT scans.
• The system aids in identifying potentially high-risk patients, facilitating early intervention.
• This technology can support medical resource optimization and proactive patient management.