Objective To explore the feasibility of contrast-enhanced CT(CECT)radiomics models based on machine learning in tumor source prediction of different liver metastatic adenocarcinomas.Methods The clinical and CECT image data of 317 cases were analyzed retrospectively,including 153 from non-gastrointestinal(25 from breast adenocarcinoma and 128 from lung adenocarcinoma)and 164 from gastrointestinal(95 from colorectal adenocarcinoma,41 from gastric adeno-carcinoma and 28 from pancreatic adenocarcinoma).The volumes of the tumors were segmented in the CECT images.The uAI research platform was used to extract radiomics features.Least absolute shrinkage and selection operator regression(LASSO)was used to select features.Combing with age and gender,SVM(support vector machine)classifiers models were built.Two radiologists predicted the metastatic tumor type on the basis of the image performance respectively.The ef-fectiveness of models was analyzed using the receiver operating characteristic curve(ROC).Delong test was used to evalu-ate models.Decision curve analysis was used to further explore the clinical utility of models.Calibration curves were used to assess predictive accuracy of models.Results Six radiomics features were obtained from triple-phase images by LASSO Regression.Area under the receiver operating characteristic curve(AUC)values of the combing radiomics model were 0.738.Combing with age and gender,the AUC,sensitivity,specificity and accuracy of clinical radiomics model were 0.833、0.740、0.804 and 0.771.The AUC values of two radiologists for the differential diagnosis were 0.643 and 0.664.The diagnos-tic effectiveness of the clinical radiomics model was higher than two radiologists reading,and the difference was statistically significant(P＜0.05).Conclusion Combing radiomics models of CECT showed good performance in liver metastases source prediction of gastrointestinal or non-gastrointestinal adenocarcinoma.The effectiveness of the SVM models was im-proved when combing with age and gender,obviously higher than radiologist reading.
NETL
NSTL
维普
万方数据
