Purpose: To mitigate the risk of spreading COVID-19, it is crucial to detect people with the disease as early as possible. As Artificial intelligence (AI) assisted film reading technology is being increasingly adopted in radiology clinics, we developed a fully automated AI-based system to specifically screen out COVID-19 patients using routine CT simulation images. The tool will report the probability of a patient having contracted COVID-19. Particularly, it is not a diagnostic tool.
Methods: The tool, which we call it AI-watchdog, consists of four components: monitoring of new CT simulation, prescreening of CT for the presence of lungs, detection of COVID-19 traits on CT images, and notification of positive patients to care team. As soon as a CT scan is found to be completed, it is copied to the AI-watchdog server. If lung is found on a CT scan, the AI component predicts the probability of COVID-19. The prediction is implemented using deep learning model convolutional neural network (CNN). It was trained and tested on three external datasets, and one internal dataset. To test generalizability, the model was tested on a cancer patient, who had scans both before and after contraction of COVID-19.
Results: The CT monitoring detects and make copies of new CT scans at a frequency of every 10 seconds. There is no discernable effect on the current clinic workflow. The prescreening takes 2 seconds. The prediction takes about 35 seconds, with accuracy of 96.7% and false positive 4.4%. The notification to care team includes 3 probabilities for COVID-19, common pneumonia, and normal. The model correctly identified all of the cancer patient scans.
Conclusion: An AI-based watchdog application offers extra layer of barrier to present spreading of COVID-19 in clinics of radiation therapy. It has proven to be effective and practical.