A deep learning-based application for COVID19 diagnosis on CT: The Imaging COVID-19 AI initiative


In the face of the ongoing global health crisis triggered by COVID-19, getting a timely and accurate diagnosis is crucial. Although RT-PCR remains the go-to diagnostic method, its drawbacks, like inconsistent sensitivity and longer waiting periods for results, highlight the need for additional diagnostic approaches. The objective of the Imaging COVID-19 AI Initiative was to explore the promise of chest imaging, especially CT scans, in diagnosing COVID-19. The core idea is to create a deep learning-powered clinical decision support system to automate COVID-19 diagnosis from chest CT images, along with a segmentation tool to measure the extent of lung involvement and assess the severity of the disease. 

The study, encompassing 20 institutions across seven European countries, utilized a custom 3D convolutional neural network for multi-class classification and segmentation tasks. From 2,802 CT scans of 2,667 unique patients, the diagnostic model showed strong sensitivity and specificity in identifying COVID-19 cases. The diagnostic model showed strong micro-average and macro-average AUC values (0.93 and 0.91, respectively) in the external test dataset, demonstrating a sensitivity of 87% and a specificity of 94% in spotting COVID-19 cases. The segmentation performance, measured by the Dice similarity coefficient (DSC), was moderate at 0.59. An imaging analysis pipeline was also framed, providing quantitative reports to the users. 

The Imaging COVID-19 AI initiative has successfully pioneered a deep learning-based clinical decision support system, illustrating a promising concurrent reading tool for clinicians.