The Value Of CT Radiomics In Identifying The Source Of Liver Metastases And Predicting The Efficacy Of Chemotherapy

Objectives:To investigate the feasibility of liver metastases radiomics models for predicting the source of hepatic metastases from gastrointestinal(GI)vs.non-gastrointestinal(non-GI)primary tumors on contrast enhanced CT(CECT).Methods:347 patients with liver metastases(180 from GI and 167 from non-GI)and abdominal CECT including arterial,portal venous,and delayed phases were divided into training(221)and validation(96)sets at a ratio of 7:3 and an independent testing set(30)which was screened by patients in the most recent and different period from the previous examination.Radiomics features of liver metastases were extracted from manually delineated volume of interests(VOIs)including tumoral(Vtc)and peritumoral(Vpt)regions on all three phases of CECT.For all 3 CECT phases and 2 different VOIs,6 single phase models were built,then single phase models were combined to build combined models.The least absolute shrinkage and selection operator(LASSO)regression was used to select radiomic features.And the optimal radiomics features were used in logistic regression models,using receiver operator characteristic curve(ROC)to evaluate the diagnostic efficiency.Delong test was used to evaluate models with the training and validation set.Correlations between radiomics features and sources of liver metastases were further explored by Point-Biserial correlation.Decision curve analysis(DCA)was performed to examine the actual clinical utility of the best model.Results:For CECT radiomics models,liver peritumoral features models more accurately predicted liver metastases compared to tumoral features models in 6 single phase models.The best single-phase model was a venous phase peritumoral VOI with 18 features.Area under the curve of ROC(AUC),sensitivity and specificity were 0.816,0.740 and 0.761,respectively in the validation set.while the best arterial phase tumoral VOI gave an AUC of 0.658 in the validation set.For the combined models,peritumoral VOI in arterial and venous phases model achieved the best prediction performance.15 features selected from arterial and venous phases gave an AUC of 0.926 in the validation set and 0.884 in the independent testing set in best combined model.Conclusion:Liver peritumoral features models more accurately predicted liver metastases on CECT compared to tumoral features models.And the best combined model was peritumoral arterial and portal venous phase combined radiomics model which can identify the source of hepatic metastases from GI versus non-GI malignancies.Objectives: To explore the feasibility of contrast enhanced CT(CECT)radiomics models based on machine learning to predict the response of colorectal liver metastases to first-line oxaliplatin-based chemotherapy.Methods: The clinical and images data of 74 colorectal liver metastases patients treated with first-line oxaliplatin-based chemotherapy(FOLFOX/XELOX)were analyzed retrospectively,including 227 liver metastases lesions.Retrospective analysis of baseline CECT scans before chemotherapy and imaging tests after 4 cycles of chemotherapy.Patients would be divided into responding group and non-responding group according to response evaluation criteria in solid tumors(RECIST 1.1).The volume of interests(VOIs)including tumors and peritumoral area(periphery within 2 mm)were segmented on arterial and venous phases images of baseline CECT scans before chemotherapy.The u AI research platform(u RP)was used to extract radiomics features.Least absolute shrinkage and selection operator(LASSO)regression was used to select features.For 2 CECT phases(arterial and portal venous phase)and 2 different VOIs(tumoral and peritumoral volume),4 single phase models were built,then single-phase models were combined to construct radiomics combined model.The effectiveness of models was analyzed using the receiver operator characteristic curve(ROC).Delong test was used to evaluate models.Results: The optimal single-phase model was the tumoral venous phase features model in 4 single phase models.16 radiomics features were obtained from tumoral venous phase images by LASSO regression.Area under the receiver operator characteristic curve(AUC)value,sensitivity,specificity and accuracy of the tumoral venous phase model were 0.867,0.807,0.775 and 0.787 in train set and 0.828,0.706,0.762 and 0.740 in the validation set,respectively.Combing tumoral and peritumoral arterial and venous phases features models,the AUC,sensitivity,specificity and accuracy of radiomics combined model were 0.929,0.844,0.831 and 0.836 in train set and 0.868,0.771,0.749 and 0.758 in the validation set,respectively.The predictive efficiency of the radiomics combined model was higher than single models,and the difference was statistically significant(P < 0.05).Conclusion: Radiomics models of CECT showed good performance in predicting the response of colorectal liver metastases to first-line oxaliplatin-based chemotherapy.The predictive efficiency of the logistic regression models was improved when single phase models were combined to construct tumoral and peritumoral arterial and venous phases radiomics combined model,radiomics combined model obviously higher than single phase models.