Preoperative Prediction Of Lymph Node Metastasis In Patients With Rectal Cancer Using A Multi-Modality Radiomics Model

AimsTo investigate the clinical value of multi-modality radiomics model,relying on computed tomography(CT),high resolution T2-weighted imaging(HR-T2WI)and diffusion weighted imaging(DWI)features,to predict lymph node metastasis in patients with rectal cancer.Materials and MethodsThis study retrospectively collected 82 patients with rectal cancer confirmed by histopathological examination postoperatively in Shandong Provincial Qianfoshan Hospital from July 2018 to August 2020.They both had pelvic contrast-enhanced CT and magnetic resonance imaging(MRI)examination within 2 weeks before surgery.Preoperative general information(including age,gender),complete imaging report[including tumor length,thickness,lymph node status,tumor location,mesorectal fascia(MRF)invasion,extramural venous invasion(EMVI)and proportion of the tumor occupied lumen circumferential(tumor_circumference_length)and the results of laboratory examination(including the levels of carcinoembryonic antigen(CEA),carbohydrate antigen 199(CA199)and carbohydrate antigen 724(carbohydrate antigen 724,CA724)]were collected.There were complete pathological reports of lymph nodes after operation(including pathological T stage and N stage).The included patients were divided randomly into a training cohort(n=57)and a testing cohort(n=27)at a ratio of 7:3.Univariate and multivariate logistic regression analysis were used to identify the clinical risk factors of age,gender,tumor length and thickness,MRI-reported lymph node status,tumor location,MRF,EMVI and tumor_circumference_length,CEA,CA199,CA724 and pathological T stage that correlated with lymph node metastasis.Upload venous CT and MRI images(including HR-T2WI and DWI)obtained from the picture archiving and communication system(PACS)to Huiyi Huiying workstation in DICOM format.The regions of interest(ROIs)were manually segmented from the largest cross-sectional tumor area on venous phase CT,HR-T2WI,and DWI at b-value of 1000 s/mm2 by two radiologists who were blinded to the histopathology results with 7 years of experience in rectal cancer imaging respectively.A total of 4227 radiomics features were extracted based on ROIs,and then the consistency of the extracted radiomics features were analyzed by inter-and intraclass correlation coefficients(ICCs).The variance threshold method,Select_K_Best and the least absolute shrinkage and selection operator(LASSO)were used step by step to reduce the dimension of the data to obtain the final radiomics features.Six support vector machine(SVM)classification models were,respectively,built based on venous phase CT features,on HR-T2WI features,on DWI features,on combined three imaging multimodal features(i.e.CT,HR-T2WI and DWI),on independent clinical risk factors,and on the combination of multiple imaging features and independent clinical risk factors(i.e.CT,HR-T2WI,DWI and independent clinical risk factors).Model performance was assessed by receiver operating characteristic(ROC)derived area under curve(AUC)analysis.Finally,nomogram and decision curve analysis(DCA)were performed to assess clinical usefulness of the combined prediction model of multimodal radiomics features and independent clinical risk factors.Univariate and multivariate logistic regression analysis using software SPSS(Version 26.0),P<0.05 was considered to be statistically significant.The software used to build SVM was Python3.6.The software used for ROC curves,nomogram,and DCA was R(Version 4.1.0).ResultsThere was no significant difference in baseline data between the training cohort and the testing cohort(P>0.05).13 clinical risk factors were analyzed by univariate and multivariate logical regression,and the MRI-reported lymph node status were selected as an independent clinical risk factor.In total,4227 radiomics features were extracted from venous phase CT,HR-T2WI and DWI.Based on the threshold set for ICC analysis,two radiomics features with poor consistency were eliminated(<0.75),which were wavelet-LLL_ngtdm_Busyness features of HR-T2WI and DWI,respectively.After dimension reduction,nine,ten,five and eleven remaining features were selected for CT,HR-T2WI,DWI and multi-modality(i.e.CT,HR-T2WI and DWI)respectively.The radiomics features derived from multi-modality(i.e.,CT,HR-T2WI and DWI)with an AUC of 0.782 in testing cohort,were significantly associated with lymph node metastasis and showed more robust predictive performance than from single sequence alone(AUCs:0.724 for venous phase CT,0.705 for HR-T2WI and 0.673 for DWI).Compared with MRI-reported lymph node model(AUC:0.789),comparable prediction of lymph node metastasis was found with multi-modality radiomics model.Multi-modality radiomics features were further incorporated with MRI-reported lymph node status showed the highest discrimination ability for predicting lymph node metastasis(AUC:0.872,accuracy:0.840,sensitivity:0.846,specificity:0.833).DCA showed that when the threshold probability was more than 7%,the clinical-multimodal radiomics nomogram based on this combined model provided a better net benefit than the MRI-reported lymph node prediction model in predicting lymph node metastasis.Conclusions1.The multi-modality radiomics model has the potential to predict lymph node metastasis of rectal cancer before surgery.2.The multi-modality radiomics model has similar diagnostic efficacy to the MRI-reported lymph node prediction model.3.The combined model of multimodal radiomics features and clinical independent risk factors has the highest diagnostic efficacy,and clinical-mutimodal radiomics nomogram based on the combined model can be used as a clinical tool for noninvasive prediction of lymph node metastasis in patients with rectal cancer before surgery,and can guide the individual treatment strategy of rectal cancer patients.