Correlation Of CT Texture Analysis And Pathological Differentiation Degrees In HCC

Objective:To study the correlation of texture analysis based on contrast-enhanced CT and the pathological differentiation degrees in hepatocellular carcinoma.Methods:331 patients with clinicopathologically confirmed HCC were retrospectively analyzed from January 2007 to April 2010.According to Edmondson-steiner standard,patients were divided into low differentiation group and high differentiation group,and patients in each group were randomly divided into training group and test group at a ratio of 7:3.All patients underwent upper abdominal Plain and enhanced spiral CT scans before surgery.396 texture features were extracted from each tumor 3D-ROI including Form factor,Histogram,gray-level co-occurrence matrix(GLCM),gray-level run-length matrix(GLRLM),gray level size zone matrix(GLSZM).Lasso regression model was used for CT texture features dimension reduction.Akaike information criteria(AIC)was used to develop radiomic signature.Radiomic scores were calculated for each patient.Logistic regression analysis was used to analyze the relationship between radiomic signature,clinical features,radiomic signature combined with clinical features with the degree of pathological differentiation of HCC.The performance of each logistic regression model was evaluated with the receiver operating characteristic curve(ROC),establishing corresponding logistics regression prediction models.ROC cures were then generated in the training group and the test group,and area under curve(AUC),accuracy,specificity and sensitivity of ROC were calculated.Results: 1.Lasso regression model was used for CT texture features dimension reduction.3,10,10,and 5 texture features are automatically selected in the unenhanced,arterial phase,portal phase and lag phases CT scans respectively.AIC was used to develop radiomic signature and radiomic scores were calculated for each patient.There were significant statistical differences in Rad-Score between the poor-differentiated group and the well-differentiated group in training group and test group,and the Rad-score of the poor-differentiation group was significantly higher than that of the well-differentiation group.2.The radiomic signatures of unenhanced,arterial phase,portal phase and lag phases had a certain predictive efficacy for the degree of pathological differentiation of liver cancer(AUC > 0.6),among which the portal phase had the best predictive efficacy.In the training group,the AUC,accuracy,sensitivity,and specificity of radiomic signature in the portal phase were 0.744(95%CI 0.679~0.810),0.632,0.547,and 0.819,respectively,and the AUC in the test group was 0.71(95%CI 0.605~0.815),with accuracy of 0.69,sensitivity of 0.739 and specificity of 0.581.In addition,the AUC of radiomic signature in the arterial phase was slightly higher than that in the plain and delayed phases.3.In clinical characteristics,there are only the AFP has statistically significant differences(P < 0.001;0.012)between well and poor differentiation in the training group and test group.The AUC of the logistic prediction model was 0.716 in the training group and 0.636 in the test group.The AUC of the clinical characteristics in the test group was lower than that in the portal phase and arterial phase radiomic signatures,but slightly higher than that in the plain scan phase and delayed phase.4.In the logistic regression model of radiomic signature combined with clinical features,only Rad-Score and AFP level showed statistical significance in classification of liver cancer pathological differentiation(p < 0.01).The predictive efficacy of radiomic signature combined with clinical characteristics was higher than that of radiomic signature and clinical characteristics alone.The AUC in the portal phase of the training group was 0.823,while that in the test group was 0.72,which could better predict the degree of pathological differentiation of liver cancer.Conclusion: 1.It is feasible to use CT texture analysis to predict the degree of pathological differentiation of HCC.The prediction efficiency of CT texture analysis in the portal phase is superior to CTTA in the plain scan phase,arterial phase and delayed phase and clinical features.2.The combination of radiomic signature and AFP level can more effectively predict the degree of pathological differentiation of HCC.