Automatic And Accurate Pancreas Segmentation Using Deep Learning

In recent years,the incidence of pancreatic diseases has increased significantly year by year,especially pancreatic cancer,which is one of the most lethal tumors.It is known as the "king of cancer",and its five-year survival rate is only less than 1%.With the tremendous progress of medical equipment,the rapid development of computer image processing and imaging omics technology,early accurate diagnosis of pancreatic lesions based on medical imaging big data analysis is an important way and the latest method to improve pancreatic cancer survival rate.The method for diagnosing pancreatic diseases can improve the early diagnosis rate of pancreatic diseases by clinicians,so as to implement more efficient and accurate treatment methods.However,how to implement effective,rapid and accurate pancreas segmentation is a huge problem in the analysis of medical image big data.The thesis focuses on the characteristics of the pancreas with complex anatomical structure,large morphology and position variation,and a large number of adjacent tissues around it.Based on the deep convolutional neural network,we conduct in-depth research on the pancreas segmentation problem.The research content and main contributions include the following aspects:First,a pancreas segmentation method based on DenseVoxNet is proposed.This method uses a 3D full convolution architecture to achieve the maximum information flow between the layers through a dense connection mechanism,thereby simplifying network training;It avoids learning redundant feature maps and improving network training efficiency by encouraging feature reuse.At the same time,adding auxiliary paths to the network structure strengthens the gradient propagation and stabilizes the learning process to achieve accurate pancreas segmentation.Second,a two-stage pancreas segmentation method based on target detection is proposed.The pancreas as a small target usually accounts for a very small proportion of the background.In order to achieve precise segmentation of the pancreas,the thesis uses a two-stage learning framework to achieve a coarse-to-fine segmentation of the pancreas.In the coarse segmentation stage,the Faster-RCNN framework commonly used in object detection is used to generate an tightened and precise localization bounding box of the pancreas,and fill the surroundings of the bounding box containing the pancreas to get a coarse segmentation result;In the fine segmentation stage,multiple segmentation models are used to finely segment the pancreatic coarse segmentation results to achieve precise pancreas segmentation.Compared with the current main method,this method not only improves the accuracy of pancreas segmentation,but also significantly optimizes the execution time.In the thesis,the two pancreatic segmentation methods are evaluated through the NIH pancreas segmentation data,and measured by the average Dice-S?rensen coefficient(DSC),reached 78.70%and 83.90%respectively.This experimental result provides sufficient support for the validation of the algorithm.