Machine Learning Based Radiomics Analysis For The Prediction Of Rectal Metachronous Liver Metastasis

Part 1 Enhanced CT and MR Radiomics of Rectal Cancer for Prediction of Metachronous Liver MetastasisPurpose:To build radiomic models using different machine learning algorithms based on enhanced venous phase CT and MR images for predicting metachronous liver metastases(MLM)of rectal cancer,and to compare the predictive performance among machine learning algorithms.Materials and methods:This study retrospectively analyzed 76 patients with rectal cancer without liver metastases at first diagnosis,including 38 patients in the group developed liver metastases within 24 months and 38 patients in the group without liver metastases more than 24 months.The baseline clinical characteristics(including age,gender,T stage,N stage,CEA and CA19-9)between the two groups were compared.Images segmentations were conducted on the enhanced CT and MR images of rectal cancer lesions.A total of 1029 radiomic features were extracted respectively.Feature selection were performed in the radiomic feature sets extracted from enhanced CT and MR images respectively,and the combining feature set with 2058 radiomic features incorporating CT and MR with the least absolute shrinkage and selection operator(LASSO)method.Five-fold cross-validation and six machine learning algorithms(decision tree,DT;gradient boosting,GB;K-nearest neighbor,KNN;logistic regression,LR;random forest,RF;support vector machine,SVM)were utilized for models constructing.The diagnostic performance of the models was evaluated by the receiver operating characteristic(ROC)curves with indicators of accuracy,sensitivity,specificity and area under the curve(AUC).Results:There were no significant differences between the MLM group and non-metastasis group in baseline clinical characteristics including age,sex,T stage,N stage,tumor biomarkers CEA and CA19-9(all P>0.05).After features selection using the LASSO method,the optimal feature set sizes,which selected from 1029 enhanced CT radiomic features,1029 enhanced MR radiomic features,and 2058 radiomic features(combined primary CT and MR radiomic features),were 1,4 and 5 respectively.In ModelCT,all six machine learning algorithms showed poor prediction performance(AUC:0.439-0.640).In ModelMR,the LR(AUC=0.750±0.137)and SVM(AUC=0.764±0.128)algorithms showed good prediction performance.In Modelcombined,the LR(AUC=0.742± 0.101)and SVM(AUC=0.718±0.069)algorithms also showed good prediction performance,but not more than ModelMR with LR and SVM algorithms.Conclusion:The ModelMR using the LR and SVM algorithm showed good prediction performance for rectal MLM.However,no matter which common machine learning algorithms were used,the ModelCT had limited predictive value for rectal MLM,and the Modelcombined failed to further increase the prediction performance of the ModelMR.Part 2 Machine Learning-based Analysis of Rectal Cancer MRI Radiomics for Prediction of Metachronous Liver MetastasisPurpose:To use machine learning-based MRI radiomics to predict metachronous liver metastases(MLM)in patients with rectal cancer.Materials and methods:This study retrospectively analyzed 108 patients with rectal cancer(54 patients in MLM group and 54 patients in non-metastases group).Statistical analyses of clinical characteristics were performed using SPSS software with t-test or the chi-square test.A P-value less than 0.05 was considered to denote a significant difference.The features selection and models development were prepared using the Anaconda3 platform with Python Scikit-Learn and Matplotlib packages.Volumes of interest(VOIs)were drawn manually to cover the whole rectal tumor on each consecutive slice.Radiomic features were extracted from VOIs of T2WI and venous phase(VP)images.Feature selection were performed in the radiomic feature sets extracted from images of T2WI and VP sequence respectively,and the combining feature set with 2058 radiomic features incorporating two sequences with the least absolute shrinkage and selection operator(LASSO)method.Five-fold cross-validation and two machine learning algorithms(support vector machine,SVM;logistic regression,LR)were utilized for predictive model construction.The diagnostic performance of the models was evaluated by ROC curves with indicators of accuracy,sensitivity,specificity and AUC,and compared by DeLong test.One hundred-round five-fold cross-validations followed for verification of the stability and reproducibility of the predicted results.Results:There were no significant differences between the MLM group and non-metastasis group in baseline clinical characteristics(P>0.05,respectively),including age,sex,T stage,N stage,CEA and CA199.A total of 1029 radiomic features were automatically extracted from VOIs from the T2WI and VP images of each patient.Five,8,and 22 optimal features were selected from 1029 T2WI,1029 VP,and 2058 combining features using LASSO method,respectively.Four-group models were constructed using the 5 T2WI features(ModelT2wI),the 8 VP features(ModelVP),the combined 13 optimal features(Modelcombined),and the 22 optimal features selected from 2058 features(Modeloptimal).In ModelVP,the LR algorithm with an AUC of 0.74(95%CI:0.57-0.75)showed significantly better performance than the SVM algorithm(AUC:0.68,95%CI:0.56-0.72)in predicting MLM(P=0.0303).Comparing the prediction performance among the four groups of models using the LR algorithm,the Modeloptimal using the LR algorithm showed the best prediction performance(P=0.0019,0.0028 and 0.0081,Delong test)with accuracy,sensitivity,specificity,and AUC of 0.80,0.76,0.83,and 0.87,respectively.The ROC curves between the 1-round cross-validation and the 100-round cross-validation were highly coincident in the LR and SVM algorithms of the four groups of models.Conclusion:Radiomics models based on baseline rectal MRI has high potential for MLM prediction,especially the Modeloptimal using LR algorithm.Moreover,except for ModelVP,the LR was not superior to the SVM algorithm for model construction.Part 3 Machine Learning Based Whole Liver CT Radiomics Analysis for Prediction of Rectal Metachronous Liver MetastasisPurpose:To build radiomic models using different machine learning algorithms based on the whole liver enhanced venous phase CT images for predicting rectal metachronous liver metastases(MLM).Materials and methods:This study retrospectively analyzed 88 patients with rectal cancer without liver metastases at first diagnosis,including 44 patients in the group developed liver metastases within 2 years and 44 patients in the group without liver metastases more than 2 years.The baseline clinical characteristics(including age,gender,T stage,N stage,CEA and CA19-9)between the two groups were compared by SPSS software.Images segmentations were conducted on the enhanced liver CT images.A total of 1029 radiomic features were extracted.Features selection was performed with the least absolute shrinkage and selection operator(LASSO)method.In this study,all samples were randomly divided into training set and validation set according to the ratio of 8:2.Five-fold cross-validation and six machine learning algorithms(logistic regression,LR;support vector machine,SVM;decision tree,DT;K-nearest neighbor,KNN;random forest,RF;multi-layer perception,MLP)were utilized for training model.The diagnostic performance of the models was evaluated by the receiver operating characteristic(ROC)curves with indicators of accuracy,sensitivity,specificity and area under the curve(AUC).Results:There were no significant differences between the MLM group and non metachronous liver metastases group in baseline clinical characteristics(P>0.05).Ten features were selected from 1029 radiomic features using LASSO method.In all these six models,the models using the LR algorithm(the accuracy,sensitivity,specificity and AUC of training set were 0.80,0.97,0.63 and 0.90±0.06 respectively;the accuracy,sensitivity,specificity and AUC of validation set were 0.67,0.33,1 and 0.74±0.14 respectively)and the SVM algorithm(the accuracy,sensitivity,specificity and AUC of training set were 0.73,0.60,0.86 and 0.83±0.10 respectively;the accuracy,sensitivity,specificity and AUC of validation set were 0.72,0.78,0.67 and 0.78±0.10 respectively)showed good MLM prediction performance.However,the models using the DT,KNN and MLP algorithms showed poor MLM prediction performance(the AUCs ranged from 0.56 to 0.69).Conclusion:The radiomic model based on the whole liver enhanced venous phase CT images showed good performance for predicting rectal MLM,especially the models using the LR and the SVM algorithms.