Pancreatic Image Segmentation And Prediction Of Postoperative Pancreatitis Based On Convolutional Neural Network

Endoscopic Retrograde Cholangiopancreatography(ERCP)is the main surgical method for the diagnosis and treatment of pancreaticobiliary duct diseases.However,it is very easy to cause acute pancreatitis after surgery.Postoperative pancreatitis after endoscopic retrograde cholangiopancreatography is a serious postoperative disease that can be fatal in severe cases.Therefore,it is very important to predict the postoperative pancreatitis in the clinic.At present,for the prediction of postoperative pancreatitis,a large number of studies use the patient’s age,gender and previous medical history to carry out multi-factor regression analysis and prediction.These studies are prone to problems such as low sensitivity,low prediction accuracy,and unstable prediction results.With the rapid development of deep learning technology,people have widely applied deep learning to all aspects of human life.Convolutional neural network is one of the best algorithms in deep learning.It can extract deep information in images and is widely used in the field of medical imaging.At present,convolutional neural networks have achieved tremendous results in image segmentation,lesion location,and diagnosis of medical images.In response to the problems in the prediction of traditional postoperative pancreatitis,this thesis proposes a method based on convolutional neural network to achieve image segmentation of pancreatic organs and the use of pancreatic structural images to predict postoperative pancreatitis before ERCP.In order to apply convolutional neural network to pancreas image segmentation and postoperative pancreatitis prediction,this thesis is divided into three parts.(1)3D U-Net network based on convolutional neural network,constructing coarse segmentation model and fine segmentation model to realize automatic segmentation of pancreatic computed tomography images;(2)AlexNet network based on convolutional neural network,To build a prediction model of postoperative pancreatitis,and use the labeled pancreatic CT images to achieve postoperative pancreatitis prediction;(3)based on the 3D U-Net network and AlexNet network of the convolutional neural network,construct a feature extraction model of pancreatic images And prediction model,using unlabeled pancreatic CT images to achieve postoperative pancreatitis prediction.The research results show that in the 389 cases of pancreas CT image segmentation,the highest Dice Similarity Coefficient(DSC)of the model segmentation is 91.75%and the lowest DSC coefficient is 50.64%;in the prediction of AlexNet based on convolutional neural network In the model,344 marked pancreatic CT images were used to predict postoperative pancreatitis,with a prediction accuracy of 87.75%,sensitivity of 86.25%,and specificity of 90.44%;in combination with the pancreas segmentation model and the prediction of postoperative pancreatitis In the model,344 unlabeled pancreatic CT images were used to predict postoperative pancreatitis.The prediction accuracy was 80.27%,sensitivity was 79.76%,and specificity was 80.29%.Combining the research of three parts,it is found that in the CT image segmentation of the pancreas,after the fusion of the coarse segmentation model and the fine segmentation model,the convolutional neural network model can greatly improve the segmentation accuracy of the pancreas,and the DSC similarity coefficient can reach more than 90%.Based on the convolutional neural network model,the patient’s pancreatic structure before ERCP can be used to predict postoperative pancreatitis.The convolutional neural network model can effectively help clinical staff to screen patients with potential postoperative pancreatitis to reduce the incidence of postoperative pancreatitis and reduce the severity of the disease.Compared with the post-operative pancreatitis prediction model using the annotated pancreatic CT image,the accuracy rate of using the pancreatic CT image prediction model without annotation has decreased.However,it can reduce the burden on the doctor and improve the stability of the prediction model.The results of this verify that,based on the convolutional neural network model,it is feasible to use the patient’s preoperative pancreatic CT images to predict whether the patient has pancreatitis,and it has high clinical significance.