3D CNN classification model for accurate diagnosis of coronavirus disease 2019 using computed tomography images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• Coronavirus disease (COVID-19) has caused a global health crisis with millions infected and hundreds of thousands of deaths.
• Early and accurate detection of COVID-19 is critical for patient care and public health.
• Computed tomography (CT) imaging is a valuable tool for diagnosing respiratory illnesses.

Purpose of the Study:

• To develop and evaluate AI models for detecting COVID-19 from CT images.
• To differentiate COVID-19 from common pneumonia (CP) and normal controls using CT data.
• To assess the impact of CT slice processing and superimposition depth on model performance.

Main Methods:

• Utilized a large database of 1112 patient CT scans from the China Consortium of Chest CT Image Investigation (CC-CCII).
• Investigated multiple AI solutions for COVID-19 detection and classification.
• Compared model performance on complete and segmented CT slices, analyzing CT-superimposition depths.

Main Results:

• The optimal model achieved 99.76% accuracy, 99.96% recall, and 99.35% precision for identifying COVID-19 positive CT slices.
• The three-way classification (COVID-19, CP, normal) reached 99.24% accuracy with a macro-AUROC of 0.9998.
• The study reports the highest accuracy and recall on the largest publicly available COVID-19 CT dataset to date.

Conclusions:

• Developed AI models demonstrate high efficacy in detecting COVID-19 from CT scans.
• The findings suggest a powerful tool for assisting radiologists and physicians in rapid diagnosis.
• This technology can significantly support healthcare systems, particularly during periods of high patient volume.