Diagnosis of COVID-19 Using CT image Radiomics Features: A Comprehensive Machine Learning Study Involving 26,307 Patients
Abstract
Purpose
To derive and validate an effective radiomics-based model for differentiation of COVID-19 pneumonia from other lung diseases using a very large cohort of patients.
Methods
We collected 19 private and 5 public datasets, accumulating to 26,307 individual patient images (15,148 COVID-19; 9,657 with other lung diseases e.g. non-COVID-19 pneumonia, lung cancer, pulmonary embolism; 1502 normal cases). Images were automatically segmented using a validated deep learning (DL) model and the results carefully reviewed. Images were first cropped into lung-only region boxes, then resized to 296×216 voxels. Voxel dimensions was resized to 1×1×1mm3 followed by 64-bin discretization. The 108 extracted features included shape, first-order histogram and texture features. Univariate analysis was first performed using simple logistic regression. The thresholds were fixed in the training set and then evaluation performed on the test set. False discovery rate (FDR) correction was applied to the p-values. Z-Score normalization was applied to all features. For multivariate analysis, features with high correlation (R2>0.99) were eliminated first using Pearson correlation. We tested 96 different machine learning strategies through cross-combining 4 feature selectors or 8 dimensionality reduction techniques with 8 classifiers. We trained and evaluated our models using 3 different datasets: 1) the entire dataset (26,307 patients: 15,148 COVID-19; 11,159 non-COVID-19); 2) excluding normal patients in non-COVID-19, and including only RT-PCR positive COVID-19 cases in the COVID-19 class (20,697 patients including 12,419 COVID-19, and 8,278 non-COVID-19)); 3) including only non-COVID-19 pneumonia patients and a random sample of COVID-19 patients (5,582 patients: 3,000 COVID-19, and 2,582 non-COVID-19) to provide balanced classes. Subsequently, each of these 3 datasets were randomly split into 70% and 30% for training and testing, respectively. All various steps, including feature preprocessing, feature selection, and classification, were performed separately in each dataset. Classification algorithms were optimized during training using grid search algorithms. The best models were chosen by a one-standard-deviation rule in 10-fold cross-validation and then were evaluated on the test sets.
Results
In dataset #1, Relief feature selection and RF classifier combination resulted in the highest performance (Area under the receiver operating characteristic curve (AUC) = 0.99, sensitivity = 0.98, specificity = 0.94, accuracy = 0.96, positive predictive value (PPV) = 0.96, and negative predicted value (NPV) = 0.96). In dataset #2, Recursive Feature Elimination (RFE) feature selection and Random Forest (RF) classifier combination resulted in the highest performance (AUC = 0.99, sensitivity = 0.98, specificity = 0.95, accuracy = 0.97, PPV = 0.96, and NPV = 0.98). In dataset #3, the ANOVA feature selection and RF classifier combination resulted in the highest performance (AUC = 0.98, sensitivity = 0.96, specificity = 0.93, accuracy = 0.94, PPV = 0.93, NPV = 0.96).
Conclusion
Radiomic features extracted from entire lung combined with machine learning algorithms can enable very effective, routine diagnosis of COVID-19 pneumonia from CT images without the use of any other diagnostic test.
Related articles
Related articles are currently not available for this article.