Machine Learning-based Radiomics Analysis Of CT And MRI For Prediction Of Hepatocellular Carcinoma Refractory To Conventional Transarterial Chemoembolization

Part 1 Machine learning-based Radiomics Analysis of CT for prediction of hepatocellular carcinoma refractory to conventional transarterial chemoembolizationObjective:To explore new imaging biomarkers for predicting hepatocellular carcinoma(HCC)refractory to conventional transarterial chemoembolization(c-TACE)using X-ray Computed Tomography(CT)radiomics based on different machine learning(ML)algorithms,and to compare the performance of these ML algorithms.Methods:The clinical data and CT images of HCC patients who received repeated c-TACE procedures in three hospitals were retrospectively analyzed.According to the inclusion and exclusion criteria,136 patients were selected and divided into training group(n=96)and validation group(n=40).Univariate and multivariate logistic regression analyses were performed to determine clinical and radiological factors associated with TACE refractoriness.The HCC lesions(regions of interest,ROIs)in non contrast(NC),arterial phase(AP),venous phase(VP)and delayed phase(DP)of pre-TACE CT images were manually segmented and 1219 radiomics features were extracted from each ROI.The dimension reduction methods were used to reduce the dimension of radiomics features extracted from NC,AP,VP and DP images.Ten fold cross-validation method and seven ML algorithms(decision tree,DT;gradient boosting,GB;k-nearest neighbor,KNN;logistic regression,LR;random forest,RF;support vector machine,SVM;neural network,NN)were utilized for models constructing.Receiver operating characteristic curve(ROC)was used to evaluate the predictive effectiveness of different models with indicators of area under the curve(AUC),sensitivity and specificity.Results:Univariate and multivariate logistic regression analyses identified maximum diameter of tumor and tumor number as the key clinical and radiological factors related to TACE refractoriness.Clinical and radiological prediction models based on the maximum diameter of tumor and tumor number failed to achieve a good prediction performance.Radiomics features of four phases(NC,AP,VP and DP)were screened by dimension reduction methods to obtain 9,16,9 and 5 radiomics features related to TACE-refractory HCC.Seven ML algorithms were used to construct seven radiomics models for each phase,and 28 radiomics models for four CT phases were constructed.Although the NN algorithm had better performance than the other six ML algorithms in both training group and validation group,the radiomics models constructed based on the all seven ML algorithms in the training group and validation group for the prediction of TACE-refractory HCC were not ideal,the prediction and generalization performance of these models were relatively poor.In addition,the combined model of clinical and radiological factors and radiomic features cannot improve the prediction efficiency.Conclusion:The radiomic models constructed by mainstream ML algorithms based on CT may have limited predictive value for TACE-refractory HCC.The combination of clinical and radiological factors and radiomic features cannot improve the prediction performance;the predictive value may also be limited.Part 2 Machine learning-based Radiomics Analysis of MRI for prediction of hepatocellular carcinoma refractory to conventional transarterial chemoembolizationObjective:To explore new imaging biomarkers for predicting hepatocellular carcinoma(HCC)refractory to conventional transarterial chemoembolization(c-TACE)using multi-sequences/phases magnetic resonance imaging(MRI)radiomics based on different machine learning(ML)algorithms,and to investigate the generalizable ML algorithm by comparing the performance of these ML algorithms.Methods:The clinical data and MRI images of HCC patients who received repeated c-TACE procedures in three hospitals were retrospectively analyzed.According to the inclusion and exclusion criteria,121 patients were selected and divided into training group(n=88)and validation group(n=33).Univariate and multivariate logistic regression analyses were performed to determine clinical and radiological factors associated with TACE refractoriness.The HCC lesions(regions of interest,ROI)in six sequences/phases[T1-weighted image(T1WI),T2-weighted image(T2WI),Diffusion-weighted imaging(DWI),Arterial phase(AP),Venous phase(VP)and Delayed phase(DP)]of pre-TACE MRI images were manually segmented,and 980 radiomics features were extracted from ROI of each MRI sequence/phase image.The dimension reduction methods were used to reduce the dimension of radiomics features extracted from six MRI sequences/phases images.Ten fold cross-validation method and seven machine learning algorithms(decision tree,DT;gradient boosting,GB;k-nearest neighbor,KNN;logistic regression,LR;random forest,RF;support vector machine,SVM;neural network,NN)were utilized for models constructing.Receiver operating characteristic curve(ROC)was used to evaluate the predictive effectiveness of different models with indicators of area under the curve(AUC),sensitivity and specificity.Results:Univariate and multivariate logistic regression analyses identified tumor number,tumor margin and total bilirubin as the key clinical and radiological factors related to TACE refractoriness.Clinical and radiological prediction models based on tumor number,tumor margin and total bilirubin failed to achieve a good prediction performance.Radiomics features of six MRI sequences/phases(T1WI,T2WI,DWI,AP,VP and DP)were screened by dimension reduction methods to obtain 19,10,6,15,11 and 15 radiomics features related to TACE-refractory HCC.Seven ML algorithms were used to construct seven radiomics models for each sequence/phase,and 42 radiomics models for six MRI sequences/phases were constructed.AP and VP were the two single phases with the highest average prediction efficiency using different ML algorithms.Four sequences/phases models(T1WI-radiomics model,T2WI-radiomics model,AP-radiomics model and VP-radiomics model)constructed by NN algorithm ranked first in prediction performance in both training group and validation group,NN algorithm demonstrated the best performance among the seven ML algorithms.The radiomics models constructed by NN algorithm in VP and AP achieved good prediction performance,the AUC values of the VP-radiomics model and AP-radiomics model in validation group were 0.839 and 0.744,respectively.In addition,the combined model of clinical and radiological factors and radiomic features can not improve the prediction efficiency.Conclusion:It was feasible to predict TACE-refractory HCC by constructing multi-sequences/phases MRI radiomic models based on different ML algorithms.Using different ML algorithms to construct MRI radiomic models,VP and AP were the two MRI single phases with the highest average prediction efficiency.The combination of radiomic features and clinical and radiological features can not improve the prediction performance.The radiomic model based on NN algorithm had better performance than the traditional LR algorithm,and was considered to be the best ML algorithm.Compared with LR algorithm,the model constructed by NN algorithm can not only better identify TACE-refractory HCC patients to facilitate personalized management of HCC patients,but also may become a better algorithm selection for future HCC related radiomics research.