Objective To construct radiomics-semantic models to predict the treatment resistance of chemoradiotherapy in brain gliomas based on MRI and clinical data of multicenter patients.Methods Among 2 108 brain gliomas patients from five medical institutions,132 patients had residual gliomas after surgery.The clinical risk factors and multimodal MRI were collected.All patients were divided into training set(n=95)and validation set(n=37).The treatment response of gliomas after standardized chemoradiotherapy were divided into resistant and non-resistant types.The semantic features of MRI were evaluated by two radiologists.Three different segmentation regions of interest(ROI)were delineated to extract radiomics features.And that three groups of radiomics models were con-structed based on different sequence MRIs.The radiomics model with the best predictive efficacy in each group was selected and combined with MRI semantic features,three radiomics-semantic models(combined models)were established.Finally,a MRI semantic model,three groups of radiomics models and three combined models were developed.Results Comparisons between the different models showed that the radiomics-semantic model based on pre-operative T2-fluid attenuated inversion recovery(FLAIR)sequence,had the best predictive efficacy,the area under the curve(AUC)in the training and validation sets were 0.866[95％confidence interval(CI)0.790-0.942]and 0.810(95％CI 0.667-0.952),respectively.The radiomics-semantic model based on postoperative T1 WI sequence performed the second best,with the AUC of the training and validation sets being 0.812(95％CI 0.726-0.898)and 0.711(95％CI 0.541-0.881),respectively.Conclusion The combined models based on MRI radiomics and semantic features are able to predict the treatment resistance of chemoradiotherapy in brain gliomas patients,and may be used as an important basis for optimizing treatment.

Objective To construct radiomics-semantic models to predict the treatment resistance of chemoradiotherapy in brain gliomas based on MRI and clinical data of multicenter patients.Methods Among 2 108 brain gliomas patients from five medical institutions,132 patients had residual gliomas after surgery.The clinical risk factors and multimodal MRI were collected.All patients were divided into training set(n=95)and validation set(n=37).The treatment response of gliomas after standardized chemoradiotherapy were divided into resistant and non-resistant types.The semantic features of MRI were evaluated by two radiologists.Three different segmentation regions of interest(ROI)were delineated to extract radiomics features.And that three groups of radiomics models were con-structed based on different sequence MRIs.The radiomics model with the best predictive efficacy in each group was selected and combined with MRI semantic features,three radiomics-semantic models(combined models)were established.Finally,a MRI semantic model,three groups of radiomics models and three combined models were developed.Results Comparisons between the different models showed that the radiomics-semantic model based on pre-operative T2-fluid attenuated inversion recovery(FLAIR)sequence,had the best predictive efficacy,the area under the curve(AUC)in the training and validation sets were 0.866[95％confidence interval(CI)0.790-0.942]and 0.810(95％CI 0.667-0.952),respectively.The radiomics-semantic model based on postoperative T1 WI sequence performed the second best,with the AUC of the training and validation sets being 0.812(95％CI 0.726-0.898)and 0.711(95％CI 0.541-0.881),respectively.Conclusion The combined models based on MRI radiomics and semantic features are able to predict the treatment resistance of chemoradiotherapy in brain gliomas patients,and may be used as an important basis for optimizing treatment.