Enhanced coronavirus disease staging with deep learning of chest computed tomography scan images

Abstract Background Early determination of whether a patient with coronavirus disease (COVID-19) is in progression or remission stage is crucial for optimal clinical management, but currently there is no rapid and effective solution. Methods A new 3D convolutional neural network coupled mixed convolution and attention mechanisms named COV-DSNet was proposed to analyze pulmonary computed tomography (CT) scans and clinical data screening from 578 COVID-19 patients, which made it capable for rapidly staging. Subsequently, the role of AI staging diagnosis in predicting disease exacerbation and infectivity was analyzed. Results 297 CT scans from 156 patients were included in the COV-DSNet analysis. Results showed that COV-DSNet achieved a mean area under the receiver operating characteristic curve (AUC-ROC) of 0.820 ± 0.030 (95% CI: 0.792–0.844). In a randomly assigned verification sets, COV-DSNet achieved an AUC-ROC of 0.864 (95% CI:0.750–0.947) with a sensitivity of 0.708 and a specificity of 0.921, which demonstrated higher specificity than manual reading under the pre-specified single-center scope and reader setting. COV-DSNet- derived stage diagnosis improved the clinical model only with APACHE II score and PaO 2 /FiO 2 for predicting severe progression within 7 days (AUC 0.978 vs. 0.908), demonstrated good calibration (bootstrap-corrected mean absolute error 0.047), also helped to avoid misjudgments about infectivity by Sars-cov-2 RT-PCR Ct value in part of COVID-19 patients. Conclusions The COV-DSNet system provides a rapid staging tool for COVID-19 patients in our single-center study, which may contribute to the selection of rational treatments. It also provides new ideas and solutions for early clinical management of other pneumonias. Clinical trial number Not applicable.