Radiomics-based model for prediction of TGF-β1 expression in head and neck squamous cell carcinoma

Abstract Purpose: TGF-β1 is an immunosuppressive gene that regulates a variety of activities relating to immune responses. However, the association between TGF-β1 expression and the survival rate of HNSCC patients remains unclear. This study is to explore that whether there is a connection between TGF-β1 expression and patients’ survival in HNSCC, and whether the TGF-β1 expression in HNSCC patients can be non-invasively predicted by CT-Based Radiomics. Materials and Methods: Transcriptional profiling data and clinical information were obtained from TCGA database, and then grouped basing on Cutoff value of TGF-β1 expression. 139 HNSCC patients (112 for training and 27 for validation) were selected basing on the completeness of enhanced arterial phase CT images. 3D Slicer software is used for image segmentation, and PyRadiomics package for extraction of radiomic features. The optimal features for establishing the corresponding gradient enhancement prediction models were obtained using mRMR_RFE algorithm and Repeat_LASSO algorithm. Conclusively, comprehensive performances of two models, including diagnostic efficacy, calibration and clinical practicability, were compared. Results: 483 patients were classified into two groups (high expression (n=333) and low expression (n=150)) basing on the cut-off of TGF-β1 expression (5.208), and then used for survival analysis. Kaplan-Meier curve showed that TGF-β1, as an independent risk factor, significantly decreased patients’ survival (p=<0.001). For construction of grdient enhancement prediction models, we respectively obtained two features-glrlm and ngtdm-and three radiation features-glrlm, first order _ 10percentile and gldm- using mRMR_RFE algorithm and Repeat_LASSO algorithm. The two established models showed strong predictive potentials in both training cohort and validation cohort. In training set, ROC curve shows that AUC of mRMR_RFE_GBM model is 0.911 and Repeat_LASSO_GBM model is 0.733. And it is statistically significant that AUC of mRMR_RFE_GBM model (0.911) is higher than Repeat_LASSO_GBM model (0.733); Likewise, in validation set, AUC of mRMR_RFE_GBM model is 0.849 and Repeat_LASSO_GBM model is 0.72. And the difference between two models in AUC value is not statistically significant (p=0.212). In addition, calibration curve shows high consistency between the predictive result and real value, and DCA diagram shows its good clinical practicability. Moreover, whether in training set or in validation set, there is no statistical difference in AUC values between mRMR_RFE_GBM model and LASSO_GBM model (p=0.443, p=0.912), indicating that the two models both fit well. Conclusion: TGF-β1 is an independent risk factor and significantly associated with poor prognosis. mRMR_RFE_GBM model and Repeat_LASSO_GBM model based on CT-Based Radiomics features can effectively and non-invasively predict TGF-β1 expression in HNSCC. Considering the efficacy of prediction, mRMR_RFE_GBM model is better for clinical application.