| PART 1 Multiparametric MRI radiomic model for the pretreatment prediction of response to neoadjuvant chemotherapy in osteosarcoma Objective:Radiomic features are potential imaging biomarkers for therapy response assessment in oncology.We evaluated the performance of the radiomics of multi-parametric magnetic resonance imaging(T1WI,T2WI,T1WI-CE),developed and validated based on a multicenter dataset adopting a radiomic strategy,for prediction of response to neoadjuvant chemotherapy in osteosarcoma.Materials and Methods:The clinical data of 87 patients in osteosarcoma confirmed by percutaneous biopsy and pathology from January 2014 to December 2019 were retrospectively collected.The clinical data including age,sex,Pathological type,lesion location,bone destruction type,maximum diameter of lesion,alkaline phosphatase(ALP)and lactate dehydrogenase(LDH).Axial T1-weighted imaging,T2-weighted imaging and contrast enhanced T1-weighted imaging were done before neoadjuvant chemotherapy for each patient.The 87 patients were divided into training group and verification group according to the ratio of 7:3.According to the postoperative pathological Huvos tumor necrosis classification,they were divided into into good responders(GRs)and poor responders(PRs).The clinical data of the patients in the training group and the validation group were compared with the patients in good responders and poor responsders.The workflow of radiomic analysis included the following:(1)Tumor segmentation and feature extraction:The segmentation of the axial T1WI,T2WI and T1WI-CE images of each osteosarcoma lesion was accomplished by an experienced radiologist using the ITK-SNAP,which was confirmed by another radiologist.The radiomic features of each osteosarcoma lesion were extracted using AK software.(2)Feature selection and radiomic model establishment:According to the different combinations of sequences,seven kinds of radiomic models based on T1 WI,T2WI,T1WI-CE,T1WI+T2WI,T1WI+T1WI-CE,T2WI+T1WI-CE,and T1WI+T2WI+T1WI-CE were established in the training group.The least absolute shrinkage and selection operator(LASSO)and multi-factor Logistic regression method were used for the selection of radiomic features.The radiomic model was built by logistic regression(LR)and support vector machine(SVM).The radiomic models were further tested on the independent validation cohort.Clinical risk factors were compared via univariate analysis;variables with P<0.05 were included in the clinical model.The combined model was built based on T1WI+T2WI or T1WI+T2WI+T1WI-CE features combined with clinical data.(3)Model evaluation:receiver operating characteristic(ROC)curve analysis was used to evaluat discriminative performance,and the AUC was used to quantify the discriminative efficacy of all models that were established.Results:1.Clinical features:A total of 87 patients were included in this retrospective study.In total,41 patients were included in good responders,and 47 patients were included in poor responders.There was a significant difference in age,sex and ALP between the good responders and poor responders(P values were 0.008,0.003,0.049,respectively).2.Feature extraction and selection:(1)There were 944 radiomics features were included in the feature reduction screening base on axis T1WI,T2WI and T1WI-CE.A final selection of features in the radiomic features was performed by LASSO regression with ten-fold cross-validation and multi-factor Logistic regression.Three radiomics features were finally screened based on T1WI,T2WI or T1 WI-CE sequence.(2)There were 1888 radiomics features were included in the feature reduction screening base on T1WI+T2WI,T1WI+T1WI-CE and T2WI+T1WI-CE.A final selection of features in the radiomic features was performed by LASSO regression with ten-fold cross-validation and multi-factor Logistic regression.The finally number of radiomics features base on T1WI+T2WI,T1WI+T1WI-CE,and T2WI+T1 WI-CE were three,two and four respectively.(3)There were 2832 radiomics features were included in the feature reduction screening base on T1WI+T2WI+T1WI-CE.A final selection of features in the radiomic features was performed by LASSO regression with ten-fold cross-validation and multi-factor Logistic regression.The finally number of radiomics features base on axis T1WI+T2WI+T1 WI-CE was three.3.Radiomic model establishment and evaluation:(1)The AUC value of the LR predictive model in the T1WI,T2WI,T1WI-CE,T1WI and T2WI,T1WI+T1WI-CE,T2WI+T1WI-CE,and T1WI+T2WI+T1WI-CE were 0.731,0.681,0.736,0.841,0.72,0.703,0.841 respectively on the validation set,the diagnostic sensitivity was 0.769,0.651,0.769,0.923,0.769,0.615,0.923 respectively,the specificity was 0.571,0.643,0.643,0.714,0.571,0.714,0.714 respectively.(2)The AUC value of the SVM predictive model in the T1WI,T2WI,T1WI-CE,T1WI+T2WI,T1WI+T1WI-CE,T2WI+T1WI-CE,and T1WI+T2WI+T1WI-CE were 0.703,0.714,0.731,0.764,0.67,0.747,0.764 respectively on the validation set,the diagnostic sensitivity was 0.923,0.615,0.846,0.846,0.462,0.692,0.846 respectively,the specificity was 0.929,0.857,0.643,0.786,0.969,0.786,0.786 respectively.(3)The AUC value of clinical,combined 1(clinical and LR-radiomics)and combined 2(clinical and SVM-radiomics)predictive model were 0.791,0.918 and 0.764 respectively.Conclusion:1.Logical regression(LR)and support vector machine(SVM)classification learning model can evaluate the chemotherapy efficacy of osteosarcomas,and the prediction efficiency of logical regression model was higher than that of support vector machine.2.There are some differences in the evaluation of chemotherapy efficacy of osteosarcomas of the extremities by seven imaging models based on different combinations of magnetic resonance scanning parameters,among which the imaging model based on T1WI+T2WI and T1WI+T2WI+CE had the highest prediction efficiency in the verification group.3.The efficacy of the model combined with clinical data and radiomics imaging features in evaluating the chemotherapeutic efficacy of osteosarcomas of extremities was better than that models consisting of either clinical or imaging features alone.PART 2 Multiparametric MRI radiomic model for the pretreatment prediction of synchronous lung metastases in osteosarcomaObjective:Lung metastasis is the most important cause of treatment failure in patients with osteosarcoma.The purpose of this study is to develop and evaluate multi-parameter(T1WI,T2WI,T1WI-CE)magnetic resonance imaging radiomic model in predicting simultaneous lung metastasis of osteosarcoma.As an adjuvant tool,it could guide therapeutic strategies for and personalized surveillance of patients.Materials and Methods:The clinical data of 78 patients in osteosarcoma of confirmed by percutaneous biopsy and pathology from January 2014 to December 2019 were retrospectively collected.The clinical data including age,sex,Pathological type,lesion location,bone destruction type,maximum diameter of lesion,alkaline phosphatase(ALP)and lactate dehydrogenase(LDH).Axial T1-weighted imaging,T2-weighted imaging and contrast enhanced T1-weighted imaging were done before therapy for each patient.The 78 patients were divided into training group and verification group according to the ratio of 7:3.According to follow-up chest CT or confirmed by pathology for simultaneous lung metastasis group and non-simultaneous lung metastasis group.The workflow of radiomic analysis included the following:(1)Tumor segmentation and feature extraction:The segmentation of the axial T1WI,T2WI and T1WI-CE images of each osteosarcoma lesion was accomplished by an experienced radiologist using the ITK-SNAP,which was confirmed by another radiologist.The radiomic features of each osteosarcoma lesion were extracted using AK software.(2)Feature selection and radiomic model establishment:According to the different combinations of sequences,seven kinds of imaging models based on T1WI,T2WI,T1WI-CE,T1WI+T2WI,T1WI+T1WI-CE,T2WI+T1WI-CE,and T1WI+T2WI+T1WI-CE were established in the training group.The least absolute shrinkage and selection operator(LASSO)and multi-factor Logistic regression method were used for the selection of radiomic features.The radiomic model was built by logistic regression(LR)and support vector machine(SVM).The radiomic models were further tested on the independent validation cohort.Clinical risk factors were compared via univariate analysis;variables with P<0.05 were included in the clinical model.The combined model was built based on T1WI+T2WI or T1WI+T2WI+CE features combined with clinical data.(3)Model evaluation:receiver operating characteristic(ROC)curve analysis was used to evaluat discriminative performance,and the AUC was used to quantify the discriminative efficacy of all models that were established.Results:1.Clinical features:A total of 78 patients were included in this retrospective study.In total,33 patients were included in synchronous lung metastases group,and 45 patients were included in non-synchronous lung metastases group.There was a significant difference in the diameter of the tumor between the synchronous lung metastasis group and non-synchronous lung metastasis group(P values were<0.001).2.Feature extraction and selection:(1)There were 944 radiomics features were included in the feature reduction screening base on axis T1WI,T2WI and T1WI-CE.A final selection of features in the radiomic features was performed by LASSO regression with ten-fold cross-validation and multi-factor Logistic regression.The finally number of radiomics features base on axis T1WI,T2WI and T1 WI-CE were two,four and five respectively.(2)There were 1888 radiomics features were included in the feature reduction screening base on T1WI+T2WI,T1WI+T1WI-CE and T2WI+T1WI-CE.A final selection of features in the radiomic features was performed by LASSO regression with ten-fold cross-validation and multi-factor Logistic regression.The finally number of radiomics features base on axis T1WI+T2WI,T1WI+T1WI-CE and T2WI+T1WI-CE were four,three and four respectively.(3)There were 2832 radiomics features were included in the feature reduction screening base on T1WI+T2WI+T1WI-CE.A final selection of features in the radiomic features was performed by LASSO regression with ten-fold cross-validation and multi-factor Logistic regression.The finally number of radiomics features base on axis T1WI+T2WI+T1WI-CE was three.3.Radiomic model establishment and evaluation:(1)The AUC value of the LR predictive model in the T1WI,T2WI,T1WI-CE,T1WI+T2WI,T1WI+T1WI-CE,T2WI+T1WI-CE,and T1WI+T2WI+T1WI-CE were 0.686,0.85,0.87,0.879,0.736,0.85,0.914 respectively on the validation set,the diagnostic sensitivity was 0.4,0.6,0.5,0.7,0.4,0.6,0.7 respectively,the specificity was 0.786,0.75,0.786,0.929,0.786,0.75,0.929 respectively.(2)The AUC value of the SVM predictive model in the T1WI,T2WI,T1WI-CE,T1WI+T2WI,T1WI+T1WI-CE,T2WI+T1WI-CE,and T1WI+T2WI+T1WI-CE were 0.629,0.829,0.771,0.879,0.643,0.829,0.929 respectively on the validation set,the diagnostic sensitivity was1.0,0.8,1.0,0.8,0.8,0.8,0.9 respectively,the specificity was 0.429,0.786,0.5,0.857,0.5,0.786,0.857 respectively.(3)The AUC value of clinical,combined 1(clinical and LR-radiomics)and combined 2(clinical and SVM-radiomics)predictive model were 0.779,0.957,0.943 respectively.Conclusion:1.Both logical regression(LR)and support vector machine(SVM)classification learning model can predict synchronous lung metastasis of osteosarcoma.The prediction efficiency of logical regression model of most sequence combinations was higher than that of support vector machine,but the SVM model of T1WI+T2WI+T1WI-CE had the highest performance.2.Among the seven radiomic models based on different combinations of magnetic resonance scanning parameters,the radiomic LR and SVM models with T2WI sequence were effective in predicting synchronous lung metastasis of osteosarcoma.3.The efficacy of the model combined with clinical data and optimal radiomic features in predicting synchronous lung metastasis of osteosarcoma was higher than that of clinical model and imaging model,which can provide a new decision-making basis for early clinical intervention of metastasis. |
PDF Full Text Request