A Preliminary Study Of Predicting Lymphatic Metastasis Of Rectal Cancer By MRI Radiomics Model Based On Machine Learning

Objective: To evaluate the preoperative diagnostic value of MRI diffusion coefficient(ADC)imaging model based on machine learning in the lymph node metastasis of rectal cancer.Materials and Methods: The clinical and imaging data of 80 patients with rectal cancer confirmed by surgery and pathology in the First Hospital of Lanzhou University from November 2017 to November 2019 were retrospectively analyzed.According to the ratio of 7 ∶ 3,56 training sets and 24 testing sets were randomly divided.The clinical and imaging data of 56 patients with rectal cancer from February 2020 to November 2021 were continuously collected.Each patient underwent rectal MRI scan and radical resection within two weeks.After exclusion,31 patients were included in the validation set.A.K.software was used to preprocess the ADC images of all patients,then ITK-SNAP is used to sketch the image,and Pyradiomics software is used to extract all the VOI features.Using Spearman correlation analysis and Ridge regression algorithm to select the imaging histogram label most correlated with lymphatic metastasis of rectal cancer to establish a simplex histogram model,the clinical characteristics and MR reports were analyzed by single factor Logistic regression,the characteristics of P < 0.1 and Radiomics label were retained to establish a clinichistogram model,five different machine learning classifiers were used to train in the training set through 5 fold cross validation,and in the testing set and validation set,the product under the work characteristic curve(AUC),accuracy,specificity,sensitivity,precision and F1-score were used to evaluate the classification effect of the classifier.Delong test was used to compare AUC between models and classifiers.Results: 1.The classification performance of each model and classifier based on ADC images is different.Bayes gaussian classifier is the classifier with better performance in the simplex model(0.901(0.816-0.985),0.667(0.500-0.833)).MLP classifier is the classifier with better performance in several classifiers in the clinicalRadiomics model(1.0(1.0-1.0),0.743(0.591-0.895)).2.By comparing the accuracy and F1 score values of the three classifiers in the training set,test set and validation set in the pure radiomics model and the clinicalradiomics model,it can be concluded that Bayes gaussian classifier is more balanced in accuracy and F1-score values in training set,test set and verification set,whether it is pure radiomics model or the clinical-radiomics model.3.The AUC of the clinic-radiomics model with only SVC classifier in the training set was significantly higher than that of the pure radiomics model(P < 0.05),but in the test set and the verification set,the P-value of SVC classifier and other classifiers was higher than 0.05 between the pure radiomics model and clinic-radiomics model.Conclusion: 1.Radiomics features based on DWI sequence ADC images can predict the lymph node metastasis of rectal cancer preoperatively.Adding clinical features in most classifiers cannot improve the accuracy of prediction.2.Compared with the other classifiers,the Bayes gaussian classifier performs well in all data sets and performs well in predicting lymph node metastasis in rectal cancer preoperatively,both in the pure radiomics model and the clinical-radiomics model.