Optimizing the radiomics-machine-learning model based on non-contrast enhanced CT for the simplified risk categorization of thymic epithelial tumors: A large cohort retrospective study

PURPOSE: This study aimed to establish and compare the radiomics machine learning (ML) models based on non-contrast enhanced computed tomography (NECT) and clinical features for predicting the simplified risk categorization of thymic epithelial tumors (TETs). EXPERIMENTAL DESIGN: A total of 509 patients with pathologically confirmed TETs from January 2009 to May 2018 were retrospectively enrolled, consisting of 238 low-risk thymoma (LRT), 232 high-risk thymoma (HRT), and 39 thymic carcinoma (TC), and were divided into training (n = 433) and testing cohorts (n = 76) according to the admission time. Volumes of interest (VOIs) covering the whole tumor were manually segmented on preoperative NECT images. A total of 1218 radiomic features were extracted from the VOIs, and 4 clinical variables were collected from the hospital database. Fourteen ML models, along with varied feature selection strategies, were used to establish triple-classification models using the radiomic features (radiomic models), while clinical-radiomic models were built after combining with the clinical variables. The diagnostic accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC) of radiologist assessment, the radiomic and clinical-radiomic models were evaluated on the testing cohort. RESULTS: The Support Vector Machine (SVM) clinical-radiomic model demonstrated the highest AUC of 0.841 (95% CI 0.820 to 0.861) on the cross-validation result and reached an AUC of 0.844 (95% CI 0.793 to 0.894) in the testing cohort. For the one-vs-rest question of LRT vs HRT + TC, the sensitivity, specificity, and accuracy reached 80.00%, 63.41%, and 71.05%, respectively. For HRT vs LRT + TC, they reached 60.53%, 78.95%, and 69.74%. For TC vs LRT + HRT they reached 33.33%, 98.63%, and 96.05%, respectively. Compared with the radiomic models, superior diagnostic efficacy was demonstrated for most clinical-radiomics models, and the AUC of the Bernoulli Naive Bayes model was significantly improved. Radiologist2's assessment achieved a higher AUC of 0.813 (95% CI: 0.756-0.8761) than other radiologists, which was slightly lower than the SVM clinical-radiomic model. Combined with other evaluation indicators, SVM, as the best ML model, demonstrated the potential of predicting the simplified risk categorization of TETs with superior predictive performance to that of radiologists' assessment. CONCLUSION: Most of the ML models are promising in predicting the simplified TETs risk categorization with superior efficacy to that of radiologists' assessment, especially the SVM models, demonstrated the integration of ML with NECT may be valuable in aiding the diagnosis and treatment planning.