A Deep Learning System to Diagnose COVID-19 Pneumonia Using Masked Lung CT Images to Avoid AI-generated COVID-19 Diagnoses that Include Data outside the Lungs

Abstract

Objective: The objective of the current study was to develop a novel, artificial intelligence (AI)-based system to diagnose coronavirus disease (COVID-19) using computed tomography (CT) slice images. Prior research has demonstrated that, if not focused on the lungs, AI diagnoses COVID-19 using information outside the lungs. The inclusion of CT training data from multiple facilities and CT models may also cause AI to diagnose COVID-19 with features that are irrelevant to COVID-19. Thus, the objective of the current study was to evaluate a combination of lung mask images and CT slice images from a single facility, using a single CT model, and use AI to differentiate COVID-19 from other types of pneumonia based solely on information related to the lungs. Method: By superimposing lung mask images on image feature output using an existing AI structure, it was possible to exclude image features other than those around the lungs. The results of this model were also compared with the slice image findings from which only the lung region was extracted. The system adopted an ensemble approach. The outputs of multiple AIs were averaged to differentiate COVID-19 cases from other types of pneumonia, based on CT slice images. Results: The system evaluated 132 scans of COVID-19 cases and 62 scans of non-COVID-19 cases taken at the single facility using a single CT model. The initial sensitivity, specificity, and accuracy of our system, using a threshold value of 0.50, was shown to be 95%, 53%, and 81%, respectively. Setting the threshold value to 0.84 adjusted the sensitivity and specificity to clinically usable values of 76% and 84%, respectively. Conclusion: The system developed in the current study was able to differentiate between pneumonia due to COVID-19 and other types of pneumonia with sufficient accuracy for use in clinical practice. This was accomplished without the inclusion of images of clinically meaningless regions and despite the application of more stringent conditions, compared to prior studies.