Patients suspected of COVID-19 infection are tested using the ‘gold standard’ RT-PCR assay – however there can be significant time delays in processing such samples, and an associated false negative rate. This can be extremely serious for vulnerable patients with underlying medical conditions and for whom the earliest possible intervention is critical. CT scans offer a quick (< 1hr) and efficient alternative means of diagnosing COVID-19, the radiographic data providing radiologists with evidence of disease etiology. COVID-19 however be difficult to discern from community acquired pneumonias, other pulmonary disorders and against the diverse background of 'normal' lung conditions - particularly in the earliest stages of infection. This is where Artificial Intelligence (AI) could play a critical role in supporting radiology analysis, by training an algorithm on a large dataset of CT scans of known status, and ensuring that all of the image to be used has been standardised across the differing imaging protocols used at CT scan facilities. We will use cutting-edge AI infrastructure to (i) standardise thousands of publicly available chest CT scans by using Generative Adversarial Networks (GAN) so as to correct for variations in imaging protocols between differing radiology facilities, (ii) re-assess the performance of two recently published open source convolutional neural network (CNN) based classifiers developed by groups in China that have been previously shown to demonstrate high sensitivity/specificity in diagnosing COVID-19 in CT scans, (iii) work with UHG clinician radiologists to characterise the full range of lesions identified in this process so as ensure expert oversight in highlighting regions-of-interest (ROI) on flagged scans and (iv) deploy the trained AI solution to a stand-alone desktop system to allow radiologists to import CT scans and have a predicted diagnostic assessment and highlighted ROI within minutes in the clinic.
