Radiomics based on

Background: Radiomics features extracted from baseline Methods: A total of 522 and 382 patients pathologically diagnosed with DLBCL were enrolled for complete regression and 2-year event-free survival prediction, respectively. Three lesion selection methods (largest or hottest lesion, patient level) and five segmentation methods (manual and four semiautomatic segmentations) were applied. A total of 112 radiomics features were extracted from the lesions and at the patient level. The feature selection was performed via random forest, and models were constructed via eXtreme Gradient Boosting. The performance of all the models was evaluated via the area under the curve (AUC), which was compared via the Delong test. Results: The AUC values varied from 0.583 to 0.768 for the treatment response and prognosis prediction models on the basis of different lesion selection and segmentation methods. However, the prediction performance gap was not significant for each model (all P>0.05). The combined models (AUC =0.908 and 0.837 for treatment response and prognosis prediction, respectively) that incorporated radiomics and clinical features exhibited significant predictive superiority over the clinical models (AUC =0.622 and 0.636, respectively) and the international prognostic index model (AUC =0.623 for prognosis prediction) (all P<0.05). Conclusions: Although there are differences in the selected radiomics features among lesion selection and segmentation methods, there is no substantial difference in the predictive power of each radiomics model. In addition, radiomics features have potential added value to clinical features.