Objective  To explore the value of using deep learning radiomics to preoperatively predict thymoma risk classification based on enhanced CT images.

  Methods  139 patients with thymoma from Yijiashan Hospital of Wannan Medical College from January 2015 to January 2023 were selected as the training set, and 81 patients from The People's Hospital of Chizhou and The Second People's Hospital of Wuhu as the external validation set. There were 137 cases in the low-risk group (types A, AB, B1) and 83 in the high-risk group (types B2, B3). Based on the images from the CT venous phase, we extracted hand-crafted radiomics (HCR) and deep learning (DL) features and constructed Radscore. We used single- and multi-factor Logistic regression analyses to screen independently influencing factors in imaging, predicting a high risk of thymoma. We used four types of machine learning modeling methods, Logistic regression (LR), random forest (RF), decision tree (DT), and support vector machine (SVM), for modeling. We selected the best model as the output model and constructed the model nomogram, calibration curve, and clinical decision curve analysis (DCA). We used the area under the receiver operating characteristic (ROC) curve (AUC) and net reclassification index (NRI) to evaluate the model performance.

  Results  The AUCs of LR, RF, DT, and SVM for constructing prediction models were 0.899, 0.835, 0.897, and 0.860 for the training set and 0.889, 0.830, 0.905, and 0.858 for the external validation set, respectively. We chose the LR model in the training set as the output model for this study owing to its outstanding performance. Calibration curves and DCA indicated that the model had a high degree of calibration and clinical applicability. The performance of the nomogram improved compared to Radscore, with an NRI training set of 7.5% (P=0.007) and an external validation set of 5.3% (P=0.020).

  Conclusions   Based on the radiomics of the enhanced CT deep learning imaging panel, accurate risk classification of thymomas has been achieved. The nomogram can provide personalized prediction outcomes.