CT Texture Analysis Of Synchronous Colorectal Liver Metastases May Predict Response After Chemotherapy:Initial Results

PurposeThe aims of the present study are:1.To explore the correlation between CT texture features and the early chemotherapy response in patients with Colorectal Liver Metastases(CRLM).2.To establish a series of predictive classification model of CT texture features and CT texture-clinical features in order to distinguish CRLM patients between responder group and non-responder group.Materials and MethodsSixty-two patients with colorectal liver metastases were retrospectively included in the study.Chemotherapy response was classified by two radiologists according to RECIST-1.1 criteria,and then patients with different response were the classified as responder group or non-responder group.The ITK-SNAP software was used to delineate the region of interest on the CT images of the portal phase before chemotherapy,and the scientific analysis software Spyder was used for image filtering,feature extraction and model construction.Random forest(RF)Gini index method was used to score the importance of variables.The first twenty texture features were selected to construct a predictive classification model using K-Nearest Neighbor(KNN)algorithm and RF algorithm respectively.Then a CT texture-clinical features model was constructed by combining the twenty texture features and the clinical indicators that may affect the treatment response.The differences of continuous variables between different groups were tested by independent sample t-test,analysis of variance or Mann-Whitney U test,Kruskal-Wallis H test,and the categorical variables were tested by chi-square test or Fisher accurate test.Result1.General case dataA total of 28 cases were included in the responder group,including 0 cases of Complete Response(CR)and 28 cases of Partial Response(PR),and 34 cases of nonresponder group,including 26 cases of Stable Disease(SD)and 8 cases of Progressive Disease(PD).There was significant difference in age among different chemotherapy response groups(P=0.030),but there was no significant difference in age between responder group or nonresponder group(P>0.05).And there was no significant difference in gender,location of primary tumor,number of metastatic tumor,maximum diameter of metastatic tumor,targeted therapy and CEA level among different groups(P>0.05).2.Comparison of CT texture featuresThere was no significant difference in CT texture features between responder group or nonresponder group.However,the unfiltered uniformity(U)was the only CT texture features found to be significant difference in different chemotherapy response groups(P=0.013).The unfiltered U in PD group was slightly higher than that in PR group(P=0.031)and SD group(P=0.012),which was no significant difference between PR group and SD group(P=1.000).3.CT texture feature selection and prediction model constructionThe classification efficiency of the two predictive model based on KNN algorithm is not good.The average accuracy of the predictive classification model of CT texture features based on KNN algorithm is 62.6%,the AUC is 0.656(95%CI:0.60,0.71).And the average accuracy of the predictive classification model of CT texture-clinical features based on KNN algorithm is 67.0%,and the AUC is 0.658(95%CI:0.60,0.70).Compared with the predictive classification model based on KNN algorithm,the predictive classification model based on RF algorithm has good classification efficiency.The average accuracy of the CT texture features prediction model based on RF algorithm is 65.9%,and the AUC is 0.784(95%Cl:0.75,0.80).Its classification performance is better than the CT texture feature-clinical features prediction model based on RF algorithm.The average accuracy of the CT texture feature-clinical features prediction model based on RF algorithm is 66.0%,and the AUC is 0.717(95%CI:0.67,0.75).ConclusionsCT texture features(especially uniformity)are potential imaging biological indicators to predict the early chemotherapy response of CRLM.A classification model based on CT texture features has a good effect on predicting early chemotherapy response of colorectal liver metastases.