MRI-based Radiomics For Prediction Of Neoadjuvant Therapy Outcomes In Rectal Cancer

Objective: Rectal cancer is one of the most common gastrointestinal malignancies in the world.Accurate prediction of neoadjuvant therapy for rectal cancer patients is the key to achieve personalized therapy.In view of the clinical demand for predicting the efficacy of neoadjuvant therapy for rectal cancer before treatment,this study constructed an MRI based prediction model from four aspects.Materials and Methods: Approved by the Ethics Committee of Hubei Cancer Hospital,74 patients with rectal cancer who received complete neoadjuvant therapy in hubei Cancer Hospital from January 2014 to December 2020 were retrospectively collected,including 23 patients in the effective treatment group and 51 patients in the ineffective treatment group.Three experienced radiologists used ITK-SNAP open source software to delineate the lesions on the small-field T2 WI,DWI,and T1 WI enhancement sequences layer by layer.Finally,all images were reviewed by a radiologist with more than 10 working experience.In feature extraction,Pyradiomics package was used to extract 1561 self-defined image omics features.1000 deep learning features were extracted using Inception Res Net V2,VGG19,Inception V3,Res Net50,Xception network models pretrained on Image Net dataset.In feature selection and dimension reduction,single-factor logistic regression,t test and rank sum test,GBDT,correlation analysis and other methods are used.In terms of modeling,LR,SVM,NB,KNN,DT and RF methods are used for modeling.In terms of model evaluation,accuracy,sensitivity,accuracy,F1 score,AUC value,ROC curve and other evaluation indexes were used to evaluate the model.Results: The 74 collected data were randomly divided into training sets and test sets according to 7:3 for model establishment and evaluation.1.In terms of single-sequence and multi-sequence prediction performance research,1561 3D image omics features were extracted by Py Radiomics package.SVM classifier was used for modeling through three-step feature selection of singlefactor logistic regression,GBDT and correlation analysis.The AUC and Accuracy of T1 WI enhanced sequence on the test set were 0.625 and 0.594,respectively.The AUC and Accuracy of small-field T2 WI sequences on the test set were 0.672 and 0.656,respectively.The AUC and Accuracy of ADC sequences on the test set were 0.684 and0.688,respectively.The AUC and Accuracy of DWI sequences on the test set were0.891 and 0.812,respectively.The AUC and Accuracy of T1 WI enhanced sequence and DWI fusion sequence on the test set were 0.672 and 0.594,respectively.The AUC and Accuracy of the fusion sequence of small field T2 WI and DWI sequences on the test set were 0.75 and 0.68,respectively.The AUC and Accuracy of the fusion sequence of ADC sequence and DWI sequence in the test set were 0.777 and 0.75,respectively.2.In terms of different delineation methods and different prediction models,1561 image omics features were extracted from DWI sequences using Py Radiomics package,dimensionality reduction was carried out through single-factor logistic regression,GBDT and correlation analysis,and modeling analysis was carried out through six classifiers.In the LR model,The AUC of 2D,pseudo 3D and 3D on the test set was 0.785,0.836 and 0.625,respectively;Accuracy was 0.781,0.75 and 0.594,respectively.In the SVM model,the AUC of 2D,pseudo 3D and 3D on the test set was0.789,0.801 and 0.891 respectively,and the Accuracy was 0.75,0.781 and 0.812 respectively.In the NB model,the AUC of 2D,pseudo 3D and 3D on the test set was0.758,0.719 and 0.658 respectively,and the Accuracy was 0.719,0.688 and 0.562 respectively.In KNN model,the AUC of 2D,pseudo 3D and 3D on the test set is0.803,0.76 and 0.688 respectively,and the Accuracy is 0.719,0.688 and 0.562 respectively.In the DT model,the AUC of 2D,pseudo 3D and 3D on the test set is0.725,0.688 and 0.637,and the Accuracy is 0.655,0.688 and 0.688,respectively.In the RF model,the AUC of 2D,pseudo 3D and 3D on the test set was 0.809,0.91 and0.715 respectively,and the Accuracy was 0.719,0.906 and 0.688 respectively.3.Research on deep learning features extracted from different network models The Inception Res Net V2,VGG19,Inception V3,Res Net50,Xception network models pretrained on Image Net dataset were used to extract 1000 deep learning features from DWI sequence pseudo-3D sketch data.T test,rank-sum test and correlation analysis were used for feature dimension reduction.SVM classifier was used for modeling.In the test set,the AUC of Inception Res Net,VGG19,Inception V3,Res Net50 and Xception were 0.715,0.586,0.719,0.914 and 0.766,respectively.Accuracy were 0.656,0.562,0.625,0.812 and 0.688,respectively.4.In terms of fusion feature research,1000 deep learning features and 1561 image omics features were extracted from pseudo 3D drawing data of DWI sequence respectively,and dimensionality reduction was carried out by single-factor logistic regression,GBDT,correlation analysis,and six classifiers were used for modeling.In RF classifier,the AUC and Accuracy of deep learning feature on test set are 0.759 and0.625 respectively.The AUC and Accuracy were 0.91 and 0.906,respectively.The AUC and Accuracy of fusion feature on the test set were 0.959 and 0.938,respectively.Conclusion: 1.Compared with the other three sequences,DWI sequence has the best prediction performance.It is necessary to select the optimal sequence through fusion sequence research.2.Different delineation methods and different modeling methods can predict the efficacy of neoadjuvant therapy for rectal cancer.3.Compared with the deep learning features extracted from the other four network models,Res Net50 network model has the best predictive performance.4.The fusion feature can improve the prediction performance of the model.