Study Of Extramural Venous Invasion Diagnostic Model For Rectal Cancer Based On High-resolution MRI Radiomics Method

Objective:To investigate the diagnostic value of the extramural venous invasion radiomics model of rectal cancer based on high-resolution MRI.Methods:Retrospective analysis of 275 patients with postoperative pathologically confirmed rectal adenocarcinoma who had undergone rectal high-resolution MRI plain and enhanced examinations at the First Hospital of Jilin University from January2016 to August 2019.The patients were divided into cross-validation set（192 patients）and test set（83 patients）using EMVI evaluated by senior physicians as the gold standard.Clinical information and imaging semantic information of patients were collected,and clinical feature models were developed based on statistically significant（p<0.05）features.The tumor primary foci were outlined on the T2 WI using ITKSNAP,and the outer edge of the intestinal wall was outlined on the T2 WI and contrast enhanced T1WI（CE-T1WI）at the level where the tumor was located,and the Ring ROI was formed by post-processing with 3 mm of inward and outward expansion each.The ROIs formed by the above process are referred to as T2 tumor,T2 ring and enhancing ring,respectively.In this study,five radiomics models were established,namely T2 tumor（R1 model）,T2 ring（R2 model）,T2 tumor + T2 ring（R3 model）,T2 tumor + T2 ring + enhancing ring（R4 model）and T2 ring + enhancing ring（R5model）.After normalizing the original image data,the RIAS software was applied to extract the radiomics features of the outlined ROIs,and LASSO was applied to reduce the dimensionality of the extracted multiple features to screen the relatively more meaningful features,and the screened features were used to build 3 classifiers,RFC,SVM and LR.The ratio of the cross-validation and test sets was 7:3,and the ROC curves were plotted and the AUC values were calculated by using the five-fold crossvalidation method.The effectiveness of the three classifiers was compared,and the best diagnostic model was selected from them,and the sensitivity and specificity indexes were calculated to judge the effectiveness of the classifiers.Five radiomics models were combined with clinical features to remodel and evaluate the effectiveness of the combined model.Results:Patients’ gender,CEA,CA19-9,depth of tumor infiltration,MRI T-stage,MRIreported lymph node status,and MRF status were statistically significant（p<0.05）.Clinical characteristics model was developed based on the above meaningful clinical characteristics,with the AUCs of 0.813 and 0.831 in the training set and the test set respectively,sensitivity of 0.33 and specificity of 0.97 in the test set.Among the radiomics models,the R5 model had the best performance,with the AUCs of 0.959 and 0.915 in the training set and the test set respectively,sensitivity of 0.89 and specificity of 0.75 in the test set according to the Jorden index.Among the radiomics joint clinical features models,the R5 joint clinical features model（R5-C model）had the best performance,with the AUCs of 1.00 and 0.900 in the training set and the test set respectively,sensitivity of 0.61 and specificity of 0.92 in the test set,but its efficacy was still lower than that of the R5 model,which could be increased by combining clinical features with other radiomics models.Conclusion:The R5 model diagnoses EMVI with optimal efficacy,providing an effective aid for clinical decision making,better realizing individualized precision medicine.