Development And Validation Of A Novel Multimodal Radiomics Model For Preoperative Prediction Of Lymphovascular Invasion In Rectal Cancer

Objective:To explore a new predictive model of lymphatic vascular infiltration(LVI)in rectal cancer based on magnetic resonance(MR)and computed tomography(CT).Methods:A retrospective study was conducted on 94 patients with rectal cancer(46 positive and 48 negative)who underwent MR and CT examinations,obtained surgical pathological confirmation and lymphovascular invasion pathological results within 2 weeks in the first hospital of Jilin University from June 2016 to October 2018.They were randomly divided into the training cohort(n = 65)and the validation cohort(n = 29)at a ratio of 7:3.On each slice of the tumor,we delineated the region of interest on T2-weighted images,diffusion weighted images,and portal venous-phase CT images,respectively.Feature extraction was conducted in A.K.software.The student t-test or Mann-Whitney U test,Spearman’s rank correlation and least absolute shrinkage and selection operator(LASSO)algorithm were used successively to select the optimal features.Radiomics score was constructed based on multivariate logistic regression.We used two methods to establish the multimodal radiomic model for predicting LVI.The method A was to select the features in the three single modes respectively,using logistic regression to form three Rad-scores(CT_Score,T2WI_Score and DWI_Score),and then combine the three scores to generate a new Rad-scoreof multimodal model.The method B was integrate the three modal features and select the features uniformly,to generate the Rad-score of multimodal model.Finally,a receiver operating characteristic(ROC)curve was drawn.the AUC,specificity,sensitivity and accuracy of the two multimodal models and the single modal model are calculated to determine their predictive efficacy for LVI in the training and validation cohorts.We use the Hosmer-Lemeshow test to assess the goodness-of-fit and use a decision curve analysis(DCA)to evaluate the clinical usefulness of the multimodal radiomics via calculating the net benefit at different threshold probabilities.Finally,a nomogram from the best model was conducted.Results:The age,gender,CEA,CA199,tumor thickness and length between positive and negative groups were no significant difference.A.K.software extracted 396 radiomic features from each mode,a total of 1,188 features from a patient image.After feature selection and establish the model,the optimal model is model A.The model A showed that the area under the curve(AUC),specificity,sensitivity and accuracy in the training and validation cohorts were 0.884(95%CI: 0.803-0.964),0.727,0.938,0.831 and 0.876(95%CI: 0.721-1.000),0.800,0.929,0.862 respectively.The Hosmer-lemeshow test showed that P value exceeded 0.05.Calibration curve suggested that the model was fitted better,and DCA showed that the multimodal radiomics model provides greater clinical benefits.Conclusion:Multimodal(MR/CT) radiomics models can serve as an effective visual predictive tool for LVI in rectal cancer.