A novel CT-based radiomics in the distinction of severity of coronavirus disease 2019 (COVID-19) pneumonia

Abstract

Background: Convenient and precise assessment of the severity in coronavirus disease 2019 (COVID-19) contributes to the timely patient treatment and prognosis improvement. We aimed to evaluate the ability of CT-based radiomics nomogram in discriminating the severity of patients with COVID-19 Pneumonia.

Methods: A total of 150 patients (training cohort n = 105; test cohort n = 45) with COVID-19 confirmed by reverse transcription polymerase chain reaction (RT-PCR) test were enrolled. Two feature selection methods, Max-Relevance and Min-Redundancy (mRMR) and least absolute shrinkage and selection operator (LASSO), were used to extract features from CT images and construct model. A total of 30 radiomic features were finally retained. Rad-score was calculated by summing the selected features weighted by their coefficients. The radiomics nomogram incorporating clinical-radiological features was eventually constructed by multivariate regression analysis. Nomogram, calibration, and decision-curve analysis were all assessed.

Results: In both cohorts, 40 patients with COVID-19 pneumonia were severe and 110 patients were non-severe. By combining the 30 radiomic features extracted from CT images, the radiomics signature showed high discrimination between severe and non-severe patients in the training set [Area Under the Curve (AUC), 0.857; 95% confidence interval (CI), 0.775-0.918] and the test set (AUC, 0.867; 95% CI, 0.732-949). The final combined model that integrated age, comorbidity, CT scores, number of lesions, ground glass opacity (GGO) with consolidation, and radiomics signature, improved the AUC to 0.952 in the training cohort and 0.98 in the test cohort. The nomogram based on the combined model similarly exhibited excellent discrimination performance in both training and test cohorts.

Conclusions: The developed model based on a radiomics signature derived from CT images can be a reliable marker for discriminating the severity of COVID-19 pneumonia.

Keywords: COVID-19; Nomogram; Radiomics; Tomography; X-ray computed.

Fig. 1

Flowchart of radiomics procedure in this study

Fig. 2

Thin-section CT images for severe and non-severe patients. a-c Images of a 25-year-old woman with non-severe COVID-19 pneumonia (CT score = 2) who had the symptoms of dry cough and fever. The axial, coronal and sagittal CT images all presented subpleural GGO (with craving stone sign) in the lower lobes of both lungs (white arrows). d-f Images of a 55-year-old woman with non-severe COVID-19 pneumonia (CT score = 1) who had the symptom of fever. The axial, coronal and sagittal CT images all presented GGO in the anterior segment of the upper lobe of the right lung, containing air bronchogram (white arrowheads) and vascular thickening (white arrow). g-i Images of a 52-year-old man with severe COVID-19 pneumonia (CT score = 4) who had the features of fever and comorbidity (diabetes, hypertension). The axial, coronal and sagittal CT images showed diffuse large regions of GGO with partial consolidation and interlobular septal thickening (white arrow). j-l Images of a 64-year-old man with severe COVID-19 pneumonia (CT score = 4) who had the symptoms of fever and cough. The axial, coronal and sagittal CT images showed diffuse large regions of GGO, accompanying consolidation (black arrows), and beaded air bronchogram (black arrowheads)

Fig. 3

Feature selection via the least absolute shrinkage and selection operator (LASSO) binary logistic regression model. aThe LASSO regression method was utilized to select radiomic features. A 10-fold cross-validation method was utilized to screen hyperparameter (λ) of the LASSO regression model and choose the model with the smallest error (λ), b LASSO coefficient profiles of the features represent vertical lines that are drawn at the value selected via 10-fold cross-validation, and the optimized hyperparameter λ was determined to be 0.00677, and 7 radiomic features were remained. c By LASSO logistic regression analysis, 7 optimal radiomic features were identified for reconstructing the prediction model

Fig. 4

Receiver operating characteristic (ROC) curves of Radiomic features for the training (a) and test cohorts (b). The AUC for the training cohort and the test cohort was 0.95 and 0.92, respectively

Fig. 5

a Radiomics nomogram for identifying severity of COVID-19. b calibration curves of the radiomics nomogram in the training set and test cohort. The calibration curves represented calibration of the nomogram on the basis of fitting the predicted probabilities and observed probabilities. The 45° line uncovers the perfect discrimination and the dotted lines reveals the discriminative ability of the nomogram. The nearer the dotted line fits to the ideal line, the better the discriminative accuracy of the developed nomogram. c Decision-curve analysis for the radiomics nomogram. The y-axis and x-axis represent the net benefit and threshold probability, respectively. The horizontal black line indicates the assumption of all severe COVID-19 patients, while the green line indicates the assumption of all non-severe COVID-19 patients