Baseline Mri Radiomics Analysis For Predicting Metachronous Distant Metastasis In Patients With Rectal Cancer

Objective:To develop and validate the MRI-based model in predicting metachronous distant metastasis(MDM)in patients with rectal cancer.Materials and methods:This study retrospectively analyzed 123 cases of rectal cancer adenocarcinoma pathologically confirmed in our hospital from January 2015 to December 2018.The patients were divided into MDM group(n=47)and non-MDM group(n=76)according to the results of2-year follow-up,without distant metastasis in the initial enhanced CT of chest,abdomen and pelvis.Radiomics features were automatically extracted from images of T2-weighted image,venous phase(1min after injection of contrast agent),DWI(b=1000s/mm~2)and ADC respectively.The patients were randomly divided into training set and test set at 7:3.Pearson correlation coefficients(PCC)and multivariate analysis of variance(ANOVA)were used to filter features with the highest correlation with MDM(including clinical data,biochemical indicators,magnetic resonance signs and radiomics features of primary rectal lesion).Linear discriminant classifier(LDA)was utilized for constructing predictive model on the training set.The diagnostic performance of the model was evaluated by area under the curve,sensitivity,specificity,accuracy,positive predictive value,and negative predictive value.The stability of the model was verified by 5-fold cross validation.Results:(1)Univariate analysis showed that the baseline MRI extramural vascular invasion(EMVI)and the pre-treatment tumor marker CEA were significantly different between the MDM group and the non-MDM group(P values were 0.030,0.038,respectively),while gender,age,tumor location,tumor thickness,maximum level average ADC values,T stage,N stage,circumferential resection margin(CRM),tumor marker CA199 before treatment,and treatment methods showed no significant differences(P values ranged from 0.082 to 0.966).Multivariate binary logistic regression showed that baseline EMVI was an independent risk factor for rectal cancer with MDM(OR=2.174,P=0.049).(2)Each patient obtained 4 VOIs,and 1409radiomics features were extracted from each VOI,resulting in a total of5636 features per patient.After feature dimension-reduction,16optimal features were selected,including:(1)T2WI features,2;(2)Characteristics of venous phase,2;(3)DWI features,8;(4)ADC features,1;(5)Clinical radiological features,3(including treatment methods,pre-treatment CEA and baseline EMVI).The AUC,95%CI,sensitivity,specificity,accuracy,positive predictive value,and negative predictive value of the training set of the model constructed by the LDA classifier were 0.927,0.862～0.977,0.879,0.906,0.895,0.853,0.923,respectively,meanwhile the AUC,95%CI,sensitivity,specificity,accuracy,positive predictive value,and negative predictive value on the test set were 0.752,0.580～0.895,0.714,0.739,0.730,0.625,0.810,respectively.The average AUC of the training set and validation set on five-fold cross validation were 0.957 and 0.747,respectively,suggesting that the prediction model had good stability.Conclusion:Baseline MRI based model can be used as a noninvasive tool for early prediction of MDM in rectal cancer,facilitating clinical decision making and improving patient outcomes.