Automatic Segmentation Of Pancreatic CT Images Based On Deep Learning

The incidence of pancreatic cancer has increased significantly in recent years in China and abroad,and the diagnosis of early stage pancreatic cancer is very difficult,so the diagnosis and treatment of pancreatic cancer is still the focus of medical research.Currently,CT images are commonly used to diagnose early stage pancreatic cancer,but the subsequent manual review not only increases the workload of physicians,but also affects the accuracy of manual segmentation.It is natural to use deep learning for various tumor localization and organ segmentation to become an urgent goal to be solved.In this thesis,a deep learning model combining Transformer and convolutional neural network is designed for the task of pancreatic CT image segmentation,with the goal of automatically identifying and segmenting the complete and accurate pancreatic parts from the available CT images.The data in the article are obtained from the NIH public pancreas dataset,and the model synergistically integrates CNN and Transformer structures,aiming to achieve accurate localization of the segmentation target.To achieve this goal,the segmentation master network is trained in two phases.Specifically,in the coarse segmentation stage,the main network uses the a priori information of the spatial location of the pancreas obtained from the labels to achieve the accurate localization of the segmentation target;in the fine segmentation stage,this a priori information enables the network to perceive the spatial information of the segmentation target more accurately,so as to reduce the bounding box and increase the proportion of the pancreas region in the whole map,thus the segmentation result can effectively preserve the edge information of such a small organ as the pancreas This can effectively preserve the edge information and 3D structure of the pancreas,which can solve the problems of poor accuracy of single CNN segmentation and neglecting the spatial information of pancreatic CT images.