Application of a Machine Learning Fusion Model Based on Computed Tomography Image Omics in the Differential Diagnosis of Novel Coronavirus Pneumonia and Other Viral Pneumonia in Children

Objective: To investigate the application of the machine learning fusion model based on computed tomography (CT) image omics in the differential diagnosis of novel coronavirus pneumonia and other viral pneumonia in children. Method: A retrospective analysis was performed on the clinical and imaging data of 49 children under 18 years old who tested positive for novel coronavirus pneumonia by nucleic acid test and received chest CT scans at Shanxi Children's Hospital and Taiyuan Maternal and Child Health Hospital from December 2022 to February 2023. Additionally, the clinical and imaging data of 98 cases of viral pneumonia caused by other single viruses from January 2020 to January 2023 in Shanxi Children's Hospital were retrospectively analyzed. The imaging features of viral pneumonia were extracted from the chest CT images of the first non-contrast scan. Combined with the analysis of clinical data, the imaging model, clinical model and fusion model were established. The diagnostic performance of each model was analyzed by a receiver operating characteristic (ROC) curve, calibration curve and decision curve. Results: In the image group learning model in the training set differential diagnosis COVID-19 group and non-COVID-19 groups, the area under the ROC curve (AUC) was 0.854, sensitivity 86.1%, 75.2%, and accuracy 84.3%. In the test set differential diagnosis COVID-19 group and non-COVID-19 groups, the AUC was 0.839, sensitivity 84.6%, 72.1%, and accuracy 86.4%. In the clinical group learning model in the training set differential diagnosis COVID-19 group and non-COVID-19 groups, the AUC was 0.829, sensitivity 73.5%, 86.4%, and accuracy 75.3%. In the test set differential diagnosis COVID-19 group and non COVID-19 groups, the AUC was 0.821, sensitivity 70.4%, 75.1%, and accuracy 70.7%. In the fusion model in the training set differential diagnosis COVID-19 group and non-COVID-19 groups, the AUC was 0.878, sensitivity 73.4%, 75.4%, and accuracy 75.3%. Finally, in the test set differential diagnosis COVID-19 group and non-COVID-19 groups, the AUC was 0.865, sensitivity 78.5%, 87.5%, and accuracy 70.7%. The results showed that the fusion model of training set and test set had positive effect compared with the imaging omics model and clinical omics model. The calibration curve shows that the training set and test set fusion model can be used to predict COVID-19 probability and has a good consistency between observers; The decision curve shows that the fusion model can obtain better net income. Conclusion: A fusion model based on CT imaging machine learning can diagnose COVID-19 and differentiate it from other causes of viral pneumonia in children.