Identification Of Enchondroma And Chondrosarcoma In Long Bone Using Radiomics Features Extracted From Magnetic Resonance Images

Objectives:(1)Describe radiomics and imaging features extracted from MR images that can distinguish enchondroma from chondrosarcoma in long bone;(2)Evaluate the differentiating effects of features,to prepare for establishing a model for diagnosis of enchondroma and chondrosarcoma in long bone.Methods:We retrospectively reviewed 120 patients diagnosis of enchondroma and chondrosarcoma from Jan.2011 to May.2018.Patients include 70 of enchondroma,50 of chondrosarcoma.All complete clinical data,magnetic resonance imaging(MRI)scans and pathological data were reviewed.Radiomics features were extracted from the pretreatment diagnostic T1-weighted magnetic resonance images(MRI).Regions of interest(ROI)segmentation was performed and radiomics and imaging features was extracted.Minimum redundancy maximum correlation(mRNR)was performed for optimal feature selection.A stable model was constructed using multivariate logistic regression analysis and four-fold cross validation based on selected features and then evaluated power ot prediction by AUC.Results:In this study,492 radiomics features were extracted from the pretreatment diagnostic T1-weighted magnetic resonance images(MRI).Minimum redundancy maximum correlation(mRNR)was used to select 13 best predictable features and then model was constructed using multivariate logistic regression.We chose 4 fold cross-validation for validation.Thereinto,one-fold cross validation selected the following best features:HHH_GLRLM_SRHGE_T1、SZHGE_T1ZSN_T1、Uniformity_T1 and denth_T1;two-fold cross validation selected the following best features:HHL_GLCM_inf2h_T1、cshad_T1、LLH_GLCM_entro_T1、LLL_GLCM_corrm_T1 and HGZE_T1;three-fold cross validation selected the following best features:Variance_T1、ZSN_T1、LHL_GLRLM_SRHGE_T1、Uniformity_T1 and denth_T1;four-fold cross validation selected the following best features:SZHGE_T1、LLL_GLCM_corrm_T1、ZSN_T1、Entropy_T1 and denth_T1.The average value of AUC of cross validation was:training cohort:0.961 ±0.015(mean±SD);validation cohort:0.901±0.053(mean±SD).Conclusions:Extracting valuable radiomics features form MR imaging is viable.MR Radiomics features is capable of identification of enchondroma and chondrosarcoma in long bone.Thus MR radiomics features can provide a basis for decisions for personalized diagnosis and treatment.