Establishment And Validation Of Radiomics Model For Preoperative Prediction Of Vascular Invasion In Rectal Cancer Based On IVIM Functional Images

Objective:To explore the application value of IVIM functional map radiomics model in preoperative prediction of vascular invasion in patients with rectal cancer.Methods:83 rectal cancer patients（negative:positive=51:32）who underwent high-resolution rectal MRI（including diffusion sequence of multi b values）and performed surgery within 2 weeks were analyzed retrospectively.The proportion of 7:3 was randomly divided into training group and validation group（58:25）.We drew the tumor edge layer by layer in DWI（b=1000mm m2/s）and Standard ADC map,and extracted the VOI radiomic features by AK software.In addition,the traditional parameters of IVIM were measured in MADC software to obtain ADC、D、D*、f values.The model was established by using independent sample t test or Mann Whitney U test,Spearman rank correlation coefficient and GDBT（gradient promotion decision tree）algorithm to select the highest predictive performance.The radiomics score was constructed based on multivariate logistic regression.This paper establishes the model in two ways:the first is to model the traditional parameters of IVIM（ADC,D,D*,f）,and the second to model by using the radiomic features of IVIM functional maps(ADC_map,D_map,f_map).Results:IVIM parameters were measured by two independent radiologists（C,3 years of experience in abdominal imaging diagnosis;D,6 years of experience in abdominal imaging diagnosis）.The average value of parameters measured by two radiologists was taken as the final result of statistical analysis.The AUC,accuracy,sensitivity and specificity of IVIM were 0.713（0.609,0.805）,0.675,0.625 and 0.706.Two radiologists（A:4 years abdominal imaging diagnosis experience B:6 years abdominal imaging diagnosis experience）drew the primary tumor focus on three-dimensional level.402radiomic features were extracted from each one,and 2412 radiomic features were extracted from each patient’s image.The model with better prediction efficiency is built by D_map.The AUC,accuracy,sensitivity and specificity of DWI ROI mapping to D model are training group:0.803（0.697,0806）、0.684、0.864、0.571,validation group:0.800（0.619,0.955）、0.731、0.700、0.750.In addition,we found that there was a significant difference in f value between positive and negative lymphovascular invasion（P=0.002）.Conclusion:（1）In general,the prediction efficiency of the radiomics model is not inferior to that of the traditional IVIM parameters;（2）The radiomics model based on DMAP is the best one;（3）For the model based on the IVIM functional maps,the prediction efficiency of DWI ROI mapping model is better than that of ADC ROI mapping model.