Deep Learning Based On Computed Tomography For Predicting Pathological Grade Of Pancreatic Ductal Adenocarcinoma

Purpose: To develop and validate a deep learning model based on enhanced computed tomography images to predict the pathological grade of patients with pancreatic ductal adenocarcinoma before treatment.Methods: This study was conducted in two stages.The first stage is automatic segmentation of the image of pancreatic ductal adenocarcinoma and pancreas.The other stage is to predict the pathological grade of pancreatic ductal adenocarcinoma with the resulting images from automatic segmentation.A total of 363 patients were randomly divided into the training set and internal test set of the segmentation model in a 4:1 ratio from two public abdominal enhanced CT datasets of pancreas.A total of 322 patients with pancreatic ductal adenocarcinoma confirmed by biopsy or postoperative pathology from two hospitals were retrospectively collected.All of them were included in the external test set of the automatic segmentation model.According to strict selection criteria,123 patients were finally included in the study predicting the pathological grade of pancreatic ductal adenocarcinoma,and were randomly divided into the training set and the test set in a 7:3ratio.In this study,nn Uet(no-new-net)and residual network were used to construct the segmentation model and the model of predicting pathological grade respectively.Then,we train and test the model with the corresponding data set.DICE score was used to evaluate the accuracy of segmentation model.Sensitivity,specificity,positive predictive value,negative predictive value,F1 value,accuracy and area under the subject operating characteristic curve(AUC)were used to evaluate the performance of the prediction model.Results: The segmentation model based on nn Unet framework for automatic segmentation of pancreatic ductal adenocarcinoma mass and pancreatic contour obtained a DICE score of 0.755 in the internal test set,0.759 and 0.727 in the external test set 1 and external test set 2,respectively.The accuracy and AUC of the deep learning model for predicting the differentiation degree of pancreatic ductal adenocarcinoma based on residual network in the test set are 62.9% and 0.680 respectively.Conclusion: The deep learning model constructed in this study for automatic segmentation of pancreatic ductal adenocarcinoma and pancreatic contour and prediction of pathological grade of pancreatic ductal adenocarcinoma has achieved certain accuracy,but more studies with larger samples and more centers are needed to further improve the performance of the model constructed,so as to explore a convenient,safe and accurate method to predict the pathological grade of pancreatic duct adenocarcinoma before treatment in the future.