Development And Validation Of A Radiomics Model For Predicting The Risk Of Distant Metastasis After Surgery In Patients With Locally Advanced Rectal Cancer

Objective:To construct a radiomics signature to predict the risk of distant metastasis after surgery in patients with locally advanced rectal cancer(LARC).Methods:Magnetic resonance imaging(MRI)data of patients with LARC after radical surgery from January 2013 to March 2015 in three Chinese hospitals were collected.The endpoint of the current study was distant metastasis-free survival(DMFS),defined as the time from surgery to distant metastasis.All patients were divided into two groups,one group with neoadjuvant therapy and the other group without neoadjuvant therapy.For each group,data from two centers as the training set were used to construct a radiomics signature,and data from the third center as the validation set was used for external validation.Radiomics signature was developed as a predictor of DMFS based on 2136 radiomic features,extracted from T2-weighted image and diffusion-weighted image of MRI.The univariate Cox regression model,correlation analysis,and elastic-net penalized Cox regression model were used to select radiomic features and construct a radiomics signature in the training set.The patients were classified into high-risk or low-risk groups according to the median of the radiomics signature score.The potential association of the radiomics signature with DMFS was first evaluated in the training data set,and then validated in the validation data set by using Kaplan-Meier estimator.The discrimination performance of the radiomics signature was assessed by the index of concordance(C-index).Results:A total of 436 patients with the average age of 56.50 ± 11.37 years and a median follow-up time of 41.8 months were enrolled in this study.For all patients,the distant metastasis rate was 26.38%.For the group with neoadjuvant therapy,a radiomics signature was constructed by using 10 radiomic features selected in the training set,and the median of the radiomics signature score was 1.86.For the group without neoadjuvant therapy,a radiomics signature was constructed by adopting 11 radiomic features selected in the training set,and the median of the radiomics signature score was-0.13.For the group with neoadjuvant therapy,the 3-year DMFS rate of the 85 patients in high-risk group(radiomics signature score>1.86)was significantly lower than that of 85 patients in the low-risk group(radiomics signature score ≤ 1.86)(68.71%vs.88.10%,high-risk vs.low-risk group:hazard ratio[HR]=2.29,95%confidence interval[CI]:1.25-4.20,log-rank P=0.009)in the training set,and the 3-year DMFS rate of the 33 patients in high-risk group was significantly lower than that of 32 patients in the low-risk group(53.75%vs.87.39%,high-risk vs.low-risk group:HR=2.35,95%CI:1.09-5.11,log-rank P=0.031)in the validation set.For the group without neoadjuvant therapy,the 3-year DMFS rate of the 69 patients in high-risk group(radiomics signature score>-0.13)was significantly lower than that of 69 patients in the low-risk group(radiomics signature score ≤-0.13)(58.15%vs.91.26%,high-risk vs.low-risk group:HR=5.19,95%CI:2.69-10.02,log-rank P<0.001)in the training set,while the 3-year DMFS rate of the 49 patients in high-risk group was statistically indistinguishable from that of 14 patients in the low-risk group(81.63%vs.92.86%,high-risk vs.low-risk group:HR=3.09,95%CI:0.77-12.43,log-rank P=0.248)in the validation set.For the group with neoadjuvant therapy,the C-index of the radiomics signature in the training set was similar to that in the validation set(0.71,95%CI:0.63-0.79;0.72,95%CI:0.61-0.83,respectively).For the group without neoadjuvant therapy,the C-index of the radiomics signature in the training set was statistically indistinguishable from that in the validation set(0.79,95%CI:0.72-0.86;0.69,95%CI:0.55-0.84,respectively).Conclusion:The radiomics signature on MRI could be used to predict the risk of distant metastasis after surgery in patients with LARC.Objective:To develop a radiomics model for predicting the risk of distant metastasis after surgery in patients with LARC based on independent prognostic factors and radiomics signature.Methods:The clinical and pathological factors of the patients enrolled in this study were collected,including sex,age,pre-treatment clinical tumor stage(cT),pre-treatment clinical node stage(cN),and preoperative carcinoembryonic antigen(CEA)level,tumor location,surgical approach,surgical path,pathological tumor stage(ypT/pT),pathological node stage(ypN/pN),adjuvant chemotherapy,and adjuvant radiotherapy.The endpoint of the present study was DMFS,defined as the time from surgery to distant metastasis.Patients were divided into two groups,namely one group with neoadjuvant therapy and one group without neoadjuvant therapy.The independent prognostic factors for distant metastasis after surgery in patients with LARC were assessed by using univariate and multivariate Cox regression models.The clinical model was based on the clinicopathological factors,and the radiomics model was based on the combination of the clinicopathological factors and radiomics signatures by adopting multivariate Cox regression model respectively.Results:A total of 170 patients with the median follow-up time of 41.63 months were enrolled in the group with neoadjuvant therapy,while 138 patients with the median follow-up time of 37.77 months were enrolled in the group without neoadjuvant therapy.The 3-year DMFS rate of the group with neoadjuvant therapy was similar to that of the group without neoadjuvant therapy(78.50%vs.74.90%,P=0.782).Multivariate analyses revealed that cT stage(HR=2.11,95%CI:1.09-4.11,P=0.004),CEA level(HR=2.68,95%CI:1.36-5.28,P=0.028),and radiomics signature(HR=7.17,95%CI:3.68-13.94,P<0.001)were significant risk factors for 3-year DMFS of LARC in the group with neoadjuvant therapy,while the CEA level(HR=2.05,95%CI:1.10-4.58,P=0.026),pN stage(HR=2.74,95%CI:1.04-7.22,P=0.042),and radiomics signature(HR=5.39,95%Cl:1.41-20.63,P<0.001)were significant risk factors for 3-year DMFS of LARC in the group without neoadjuvant therapy.For the group with neoadjuvant therapy,the predictors of the clinical model included cT stage,CEA level,and ypN stage.The predictors of the radiomics model included cT stage,CEA level,and radiomics signature.For the group without neoadjuvant therapy,the predictors of the clinical model included CEA level,and pN stage.The predictors of the radiomics model Ⅰ included CEA level,pN stage,and radiomics signature.The radiomics model Ⅱ included the sex,surgical path,CEA level,pN stage,adjuvant radiotherapy,and radiomics signature.Conclusion:The cT stage,CEA level,and radiomics signature are the independent prognostic factors for DMFS of LARC,and can be used in a predictive model in the group with neoadjuvant therapy.And the CEA levels,pN stage,and radiomics signature are independent prognostic factors for DMFS of LARC,and can be used in a predictive model in the group without neoadjuvant therapy.Objective::To evaluate and validate the predictive ability of the radiomics model for predicting the risk of distant metastasis after surgery in patients with LARC.Methods:A total of 436 patients in three Chinese hospitals from January 2013 to March 2015 were enrolled.The pre-treatment MRI and clinical pathological characteristics of each patient were collected.The endpoint was DMFS,defined as the time from surgery to distant metastasis.All patients were divided into a group with neoadjuvant therapy and a group without neoadjuvant therapy.For each group,data form two centers as the training set were used to develop multivariate model,and data collected from the third center as the validation set was used for external validation.The predictive abilities of the radiomics model,the clinical model developed by our research,and the VN model recommended by the guidelines were evaluated both in the training and validation sets by the C-index,net reclassification improvement(NRI),integrated discrimination improvement(IDI),receiver operating curve(ROC),calibration curve,decision curve analysis(DCA),and clinical impact curve.Results:For the group with neoadjuvant therapy,the C-indexes of the radiomics model,clinical model and VN model were 0.78(95%confidence interval[CI]:0.71-0.85),0.69(95%CI:0.61-0.78),0.62(95%CI:0.53-0.72)in the training set,and 0.77(95%CI:0.68-0.87),0.63(95%CI:0.51-0.76),0.64(95%CI:0.52-0.76)in the validation set;compared with the clinical model and the VN model,the NRI values of the radiomics model were 0.13(95%CI:-0.05-0.32,P=0.173),0.44(95%CI:0.27-0.57,P<0.001)in the training set,and 0.29(95%CI:-0.002-0.58,P=0.051),0.25(95%CI:0.01-0.49,P=0.044)in the validation set,the IDI values of the radiomics model were 0.11(95%CI:0.01-0.21,P=0.031),0.19(95%CI:0.10-0.28,P<0.001)in the training set,and 0.13(95%CI:-0.02-0.29,P=0.093),0.21(95%CI:0.05-0.36,P=0.008)in the validation set;the AUC values of the radiomics model for predicting 3-year DMFS in the training and validation sets were 0.82(95%CI:0.75-0.89,P<0.001)and 0.82(95%CI:0.71-0.93,P<0.001),respectively;the Hosmer-Lemeshow calibration test revealed the good calibrations of the radiomics model predicting 3-year DMFS in the training and validation sets(P=0.561,0.908,respectively),and DCA demonstrated that the radiomics model was more useful than the clinical model both in the training(risk cutoff:0.04-0.58)and validation sets(risk cutoff:0.04-0.70).At the cut-off of 0.5 of the risk thresholds,the number of positive events predicted by the radiomics model were 88/1000,and 161/1000,respectively,and the number of true positive events were 67/1000,and 113/1000,respectively in the training set and the validation set.For the group without neoadjuvant therapy,the C-indexes of the radiomics model Ⅱ,radiomics model I and clinical model were 0.84(95%CI:0.78-0.90),0.82(95%CI:0.76-0.89),0.69(95%CI:0.61-0.77)in the training set,and 0.77(95%CI:0.66-0.89),0.72(95%CI:0.59-0.86),0.61(95%CI:0.46-0.77)in the validation set;compared with the radiomics model I and clinical model,the NRI values of the radiomics modelⅡ were-0.001(95%CI:-0.10-0.10,P=0.984),0.27(95%CI:0.09-0.45,P=0.003)in the training set,and 0.04(95%CI:-0.22-0.32,P=0.746),0.11(95%CI:-0.31-0.53,P=0.598)in the validation set,the IDI values of the radiomics model were 0.06(95%CI:-0.01-0.12,P=0.052),IDI 0.30(95%CI:0.20-0.48,P<0.001)in the training set,and 0.02(95%CI:-0.03-0.08,P=0.456),0.16(95%CI:-0.01-0.34,P=0.063)in the validation set;the AUC values of the radiomics model Ⅱ for predicting 3-year DMFS in the training and validation sets were 0.89(95%Cl:0.83-0.95,P<0.001)and 0.78(95%Cl:0.61-0.94,P<0.001),respectively;the Hosmer-Lemeshow calibration test revealed the good calibrations of the radiomics model Ⅱ predicting 3-year DMFS in the training and validation sets(P=0.833,0.278,respectively),and DCA demonstrated that the radiomics model Ⅱ was more useful than the clinical model both in the training(risk cutoff:0.04-1.00)and validation sets(risk cutoff:0.04-0.36).At the cut-off of 0.5 of the risk thresholds,the number of positive events predicted by the radiomics model Ⅱ were 182/1000,and 103/1000,respectively,and the number of true positive events were 127/1000,and 52/1000,respectively in the training set and the validation set.Conclusion:The radiomics model for predicting DMFS after surgery in LARC has good prediction ability in both the training and validation sets,and its prediction efficiency is better than the clinical model and the VN model recommended by the guidelines.