Prediction Of The Staging And Prognosis Of Rectal Cancer After Neoadjuvant Therapy Based On Magnetic Resonance Radiomics

Rectal cancer has high incidence and mortality rates in China and in the world.Traditional imaging examinations such as CT,MRI and PET/CT have important value in reflecting tumor morphology,size,location,invasion range,distant metastasis and lymph nodes,etc.,but they have limited value in evaluating efficacy and predicting prognosis after neoadjuvant treatment of locally advanced rectal cancer,and their value in accurate tumor evaluation at histological and molecular levels are also limited.However,radiomics,through in-depth analysis of high-throughput imaging feature data,microscopic structures and changing rules of tumor image,can obtain various potential quantitative information closely related to disease pathogenesis,development and prognosis,thus to more accurately evaluate tumor heterogeneity,invasiveness and therapy resistance.Radiomics can combine with clinical laboratory indexes and imaging findings,improve the accuracy of evaluation and aid in guidance of treatment and prediction of disease outcome.T-stage,N-stage is one of the bases for judging whether rectal cancer receives neoadjuvant therapy,and the prediction of T-stage and N-stage after neoadjuvant therapy can provide a strong basis for the next clinical non-radical surgical treatment,and also provide a certain basis for evaluating the efficacy and judging the prognosis.The prolongation of overall survival is the ultimate goal of treatment.Accurate prediction of overall survival is beneficial to improve the treatment plan and the life quality of patients.This study utilizes radiomics through clinical image data to conducte from the following four perspectives:(1)To explore the diagnostic value of pathological staging of rectal cancer based on MRI radiomics.(2)To explore the predictive value of T-stage(p T)of rectal cancer after neoadjuvant therapy based on multi-parameter model of MRI radiomics.(3)To explore the predictive value of N staging(p N)of rectal cancer after neoadjuvant therapy based on machine learning algorithm of MRI radiomics.(4)To explore the predictive value of overall survival(OS)of rectal cancer after neoadjuvant therapy based on machine learning algorithm of MRI radiomics.The results show that the machine learning model based on conventional MRI sequence images such as high-resolution T2 WI,T1WI enhanced sequence has relatively high accuracy and stability,which can predict the T/N transformation and prognosis of patients with rectal cancer after neoadjuvant therapy,and provide certain basis for clinical diagnosis and treatment plans.Part I MRI Radiomics for Preoperative Diagnosis Pathological Staging in Rectal CancerPurpose To explore the predictive value of pathological staging of rectal cancer based on MRI radiomics.Materials and Methods MRI images and pathological data of 127 patients with rectal cancer were retrospectively analyzed.The MRI staging(T/N)of rectal cancer was determined by MRI imaging and report,and compared with postoperative pathological T/N staging.ITK-snap software was used to manually draw ROIs(Regions of Interest)in rectal cancer foci layer by layer on T1 WI transverse axial enhanced images with the minimum contour method.The radiomics features were extracted by Pyradiomics software from ROI.The features with good stability were retained by the ICC(ICC≥0.75).MRMR and LASSO were used to screen the features most relevant to T and N staging,and radiomics models were constructed by LASSO.The fusion model was constructed by integrating age,gender,T/N staging(MRI staging)and radscore,and presented in the form of a quantitative nomogram.The ROC and the corresponding AUC,the calibration curve and decision curve were used to evaluate the predictive performance and clinical usefulness of the models.Results In this study,there were 127 patients with rectal cancer,including 82 males and 45 females.Postoperative pathological T staging(p T): 41 cases in T1-2 group(T1-2 cases,T2-39 cases)and 86 cases in T3-4 group(T3-57 cases,T4-29 cases).Postoperative pathological N staging(p N): 68 cases in the N0 group and 59 cases in the N1-2 group(N1-40 cases,N2-19 cases).According to the display of MRI,T staging(r T)was determined in 22 cases in the T1-2 group and 105 cases in the T3-4 group(T3-85 cases,T4-20 cases),with a diagnostic accuracy of 0.709.N staging(r T): 42 cases in the N0 group and 85 cases in the N1-2 group(N1-73 cases,N2-12 cases).The diagnostic accuracy was 0.575.A total of 1561 radiomics features were extracted from each case,and 11 features that were most strongly correlated with T/N staging were screened out.Based on these 11 features,a predictive model was constructed to form a radiomics label.Combined with the fusion model composed of age,gender and r T/r N,the AUC value of the fusion model(T stage)in the training corhort was calculated to be 0.833(95% CI: 0.749-0.918),and the accuracy,sensitivity and specificity were 0.798,0.800,0.793,respectively,in the testing corhort was 0.782(95% CI: 0.620-0.944),and the accuracy,sensitivity,and specificity were 0.790,0.808,and 0.750,respectively.The AUC value of the fusion model(N stage)was 0.812(95%CI: 0.723-0.901),the accuracy,sensitivity,and specificity were 0.753,0.976,and 0.562,respectively.The AUC value,accuracy,sensitivity,and specificity in the testing corhort was 0.736(95%CI: 0.561-0.912),0.763,0.778,0.750.The calibration curve indicated that the T/N staging fusion model had a good goodness of fit.The decision curves indicated that the fusion models were more clinically practical in predicting T/N staging than MRI staging.Conclusion It is feasible to diagnosis the T/N staging of rectal cancer based on MRI-T1 WI radiomics analysis,which can improve the accuracy of preoperative MRI staging and provide a certain basis for clinical selection of neoadjuvant therapy.Part Ⅱ Prediction of T-stage of rectal cancer after neoadjuvant therapy based on multi-parameter model of magnetic resonance radiomicsPurpose To explore the predictive value of T-stage(p T)of rectal cancer after neoadjuvant therapy based on multi-parameter model of MRI radiomics.Materials and Methods The clinical,imaging and pathological data of 171 patients with rectal cancer before neoadjuvant therapy were retrospectively analyzed.The patients were randomly divided into training set(n=137)and test set(n=34)with the ratio of 8:2.According to the postoperative pathological results,the patients were divided into low stage group(p T 0-2)and high stage group(p T 3-4).Statistical analysis was performed to screen out p T-related clinical characteristics(age,gender,laboratory indicators and MRI imaging findings).The same method was used in the first part to draw the rectal cancer tumor as the ROI on the images of T2 WI and the enhanced images of T1 WI respectively,and to extract the features from the ROIs,and retain the features with good stability.Then the LASSO method was used to screen the most relevant features of p T from T1 WI,T2WI and fusion features(T1WI+T2WI+ clinical features),respectively.The features most correlated with p T were screened from T1 WI,T2WI and fusion features(T1WI+ T2WI+ clinical features)using LASSO.The predictive p T models were constructed separately using the selected T1 WI radiomics features,T2 WI radiomics features,clinical features and fusion features by logistic regression(LR)method,and the performance of the models was evaluated using receiver operating characteristic(ROC)curves and calibration curves.Delong test was performed to compare the differences between the models.Decision curves were used to assess the clinical application value of the models.Results Among 171 patients with rectal cancer,71 cases were in the group with good efficacy and 100 cases were in the group with poor efficacy after neoadjuvant therapy.Among the clinical features,10 features were significantly different,which were used to construct a clinical model for predicting p T.A total of 1561 radiomics features were extracted from each sequence of images in each case.Sixteen and 18 features most strongly correlated with p T were screened from T1 WI and T2 WI respectively.The corresponding prediction model is constructed based on these features.Among the fusion features,the fusion model was constructed using the four clinical features including maximum lymph node short diameter(Lnmax),Extramural vascular invasion(EMVI),T stage(r T)and N stage(r N)evaluated by imaging,nine T1 WI,and twelve T2 WI radiomics features.The corresponding ROC curves were also drawn to evaluate the efficacy.In the training set,the areas under the curve(AUCs)of T1 WI radiomics,T2 WI radiomics,clinical imaging feature and fusion feature model were 0.868,0.921,0.713,0.967,respectively,and the corresponding AUCs in the test set were 0.761,0.842,0.689,0.932.Among the four models,the fusion model had the best predictive efficacy,with the accuracy of 0.891,sensitivity of 0.875 and specificity of 0.912 in the training set and 0.882,0.900 and 0.857 in the test set,respectively.The calibration curves showed that the fusion model has a good fit superiority.The decision curves showed that the fusion model was more clinically useful than the T1 WI,T2WI and clinical models in predicting the p T stage when the threshold probabilities in the test set ranged from 10% to 90%.Conclusion The clinical imaging findings,T1 WI and T2 WI radiomics features of rectal cancer can predict pathological T-stage after neoadjuvant treatment of rectal cancer,and the accuracy of predicting T-stage can be improved by the multi-parameter model of MRI radiomics.Part Ⅲ Prediction of N-stage of rectal cancer after neoadjuvant therapy based on machine learning algorithm of MRI radiomicsPurpose To explore the predictive value of N-stage(p N)after neoadjuvant therapy for rectal cancer based on machine learning algorithm of MRI radiomics.Materials and Methods The data of 171 patients with rectal cancer in the second part were analyzed.According to the postoperative pathological results,the patients were divided into N-(N0,no lymph node metastasis group)and N+(N1,N2,lymph node metastasis group).Statistical analysis was performed to screen out p N-related clinical characteristics.ITK-SNAP software was utilized to manually draw the rectal cancer lesion as ROI-1 and the mesentery around the rectum as ROI-2 layer by layer using the minimal contour method on the short-axis T2 WI images before neoadjuvant treatment.The same method as the second part was used to extract and screen the most relevant features of p N from ROI-1(T2WI),ROI-2(Meso)and fusion features(ROI-1+ROI-2+ clinical features),respectively.The T2 WI radiomics features,Meso radiomics features,clinical features and fusion features were used to construct the prediction model of p N in seven machine learning algorithms: support vector machine(SVM),K-proximity algorithm(KNN),random forest(RF),extreme random tree(ET),gradient lifting decision tree(XGBoost),Light GBM(LGBM)and logistic regression(LR).ROC curves were used to evaluate the performance of the models.Results: The surgical outcomes of 171 rectal cancer patients after neoadjuvant therapy showed that 92 cases were in the N-group and 79 cases were in the N+ group.Among the 7 models constructed by the features screened from the rectal tumor region(T2WI),the LR model had the best efficacy,with the training set AUC 0.899;accuracy 0.838,sensitivity 0.866,and specificity 0.812,corresponding to 0.656,0.714,0.583,0.783 in the test set.Among the models constructed by the features screened from the perirectal mesenteric region(Meso),the SVM model had the best efficacy with the training set AUC 0.972,accuracy 0.912,sensitivity 0.955,specificity 0.870,and correspondingly 0.721,0.657,0.917,0.522 in the test set.Among the clinical models,the LR model had the best predictive efficacy with the training set AUC 0.711,accuracy 0.662,sensitivity 0.627,specificity 0.697,corresponding to 0.768,0.771,0.833,0.773 in the test set.Among the fusion models,the LR model had the best predictive efficacy with the training set AUC 0.954,accuracy 0.904,sensitivity 0.910,specificity 0.899,corresponding to 0.866,0.800,0.917,and 0.739 in the test set.The LR model with fusion model has the best predictive efficacy.Conclusion: Radiomics analysis of tumor region and perirectal mesenteric region can predict the N stage of rectal cancer after neoadjuvant therapy.MRI radiomics combined with multiple sets of features using machine learning algorithm can improve the accuracy of the prediction model.Part Ⅳ Prediction of the prognosis of rectal cancer after neoadjuvant therapy based on machine learning algorithm of MRI radiomicsPurpose To explore the predictive value of overall survival(OS)of rectal cancer after neoadjuvant therapy based on machine learning algorithm model of MRI radiomics.Materials and Methods The preoperative clinical imaging data and postoperative follow-up data of 171 patients with rectal cancer in the second part were analyzed.ROI-1 and ROI-2 were selected and features were extracted using the same method as in the third part.Taking the survival status at the end of the follow-up as the cut-off point to screen features.ROC curves were used to evaluate the efficiency of the models,and the dominant models were selected and the corresponding Radscore(T2WI,Meso)was calculated.The clinical characteristics that were significantly different from the survival status at the end of follow-up were obtained by statistical analysis.The cox proportional hazard model was used to comprehensively analyze the value of Radscore(T2WI),Radscore(Meso)and the value of clinical features with predictive significance for OS prediction,which were presented in the form of quantitative nomogram,and the score of each patient was calculated.According to the median survival time,the patients were divided into the high-risk group with shorter survival time and the low-risk group with longer survival time.Kaplan-Meier curve was used to compare the survival time between the high and low risk groups in the training set and the test set,respectively.Log-rank hypothesis test was performed,and the consistency index(C-index)was used to evaluate the predictive ability of the model.Results In 171 rectal cancer patients,26 features were obtained from ROI-1 screening.Among the constructed models,SVM had the best efficacy,with AUC 0.948,accuracy 0.876,sensitivity 0.821,specificity 0.913 in the training set,corresponding to 0.857,0.794,0.929,and 0.700 in the test set.20 features were obtained from ROI-2 screening.LR had the best efficacy among the constructed models,with AUC 0.866,accuracy 0.810,sensitivity 0.696 and specificity 0.889 in the training set,and corresponding to 0.725,0.706,0.786 and 0.650 in the test cohort.The overall efficacy of the rectal tumor region models(T2WI)was better than that of the Meso models.Only the carcinoma node in perirectal mesenteric(Node)and enlarged lymph nodes in extra-mesenteric region(Extra MRF)were significant among the clinical imaging features.The four features,Radscore(T2WI),Radscore(Meso),Node and Extra MRF were input into the COX regression model.C-index=0.800 and C-index=0.7720 were obtained in the training set and in the test set respectively.Kaplan-meier curve analysis showed 1-year survival rate.In the COX regression model analysis,the score of each rectal cancer patient was calculated according to the signature composed of four variables,Radscore(T2WI),Radscore(Meso),Node and Extra MRF.There was a significant difference in the obtained signature scores between the high-risk group and the low-risk group divided by the median survival time(969 days)as the cut-off value.The C-index of the training set and the test set were 0.798 and 0.772,respectively(p < 0.005),indicating that the signature was associated with OS,and the prediction performance of the constructed model had moderate accuracy.Conclusion The radiomics model of rectal tumor region and perirectal mesenteric region combined with clinical imaging features has better OS efficacy in predicting the prognosis of rectal cancer after neoadjuvant treatment.High resolution T2 WI radiomics model based on MRI can be used to predict the prognosis of rectal cancer after neoadjuvant treatment.