Purpose: This study aimed to investigate a computational approach to triage the severe patients caused by coronavirus disease 2019 (COVID-19) pneumonia on CT image using radiomic signatures.
Methods: 263 cases were selected from a COVID-19 open database, and they were divided to 180 training [60 severe cases (critically ill and severe), 120 non-severe cases (regular and COVID-19 negative)] and 83 test (25 severe cases, 58 non-severe cases) cases. Two training datasets, each of which included 120 cases, were made from the 60 severe cases and 60 non-severe cases, which were randomly selected twice from the 120 non-severe cases. 270 radiomic features (histogram and texture features) were extracted from an original CT image and four wavelet filtered CT images. Then, 9 triage models were built by combining three feature selection with three machine learning methods. Significant features were chosen to construct radiomic signatures by using p-value-based selection, logistic regression and permutation importance. The three machine learning models were a logistic regression, random forest and support vector machine with the radiomic signatures constructed from a 5-fold cross validation. Finally, the model which showed the highest area under the receiver operating characteristics curve (AUC) for the test dataset was considered as a final classification model.
Results: In the validation of the training step, AUCs of the model using a logistic regression showed 0.875 with permutation importance. In the test, the highest AUC was 0.668.
Conclusion: Triage models using radiomic signatures could be useful in assessment of severity of COVID-19 pneumonia for clinical triage.