A Weakly-Supervised Framework for COVID-19 Classification and Lesion Localization From Chest CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate COVID-19 diagnosis is critical for patient management and outbreak control.
• Chest CT imaging is a key tool for diagnosing COVID-19.
• Deep learning offers potential for automated analysis of medical images.

Purpose of the Study:

• To develop and evaluate a weakly-supervised deep learning framework for automatic COVID-19 diagnosis and lesion localization using 3D CT volumes.
• To assess the performance of the model in terms of classification accuracy and lesion identification.
• To provide a rapid and efficient tool for COVID-19 detection.

Main Methods:

• A weakly-supervised deep learning framework utilizing 3D CT volumes.
• Lung segmentation using a pre-trained UNet.
• A 3D deep neural network for COVID-19 probability prediction.
• Lesion localization by combining activation regions and unsupervised connected components.

Main Results:

• The algorithm achieved 0.959 ROC AUC and 0.976 PR AUC.
• Classification accuracy of 0.901, positive predictive value of 0.840, and negative predictive value of 0.982.
• Processing time of 1.93 seconds per CT volume on a GPU.

Conclusions:

• The weakly-supervised deep learning model accurately predicts COVID-19 probability and localizes lesions on chest CT without lesion annotation.
• The developed algorithm is fast, high-performing, and beneficial for controlling SARS-CoV-2 outbreaks.
• The software is publicly available for use.