Research Of Pancreatic Tumor Segmentation Based On Convolutional Neural Networks

The pancreas is an important organ of the human digestive system,with exocrine and endocrine functions.The exocrine glands are responsible for the secretion of pancreatic juice for digestion of sugar,fat and protein;the endocrine glands are composed of pancreatic islets,which regulate blood sugar levels by secreting various hormones such as insulin.In the past 10 years,with the changes in the dietary structure of Chinese residents and the growth of the elderly population,the proportion of deaths caused by pancreatic cancer in cancer-related deaths has increased by 9%,and the incidence has increased year by year.Pancreatic cancer has become a major public health problem threatening people’s lives and health.Pancreatic cancer is extremely malignant,progresses quickly,and has a poor prognosis.The 5-year survival rate is less than 8%.Therefore,early detection and diagnosis is very important for pancreatic cancer patients.CT imaging examination is the most important examination method for diagnosing pancreatic diseases.This article aims to construct a model that uses CT data to segment the pancreas and tumors,so as to provide doctors with a computer-aided diagnosis system.The pancreas is soft,with complex anatomical structure,and there are a lot of tissues with similar density around it.The main difficulties in segmentation of pancreas and pancreatic tumors are: small targets,large individual differences in shape and location,blurred boundaries,and noisy labels.In response to these problems,this paper designs a segmentation algorithm based on convolutional neural networks to achieve precise segmentation of pancreas and pancreatic tumors.The main research contents are as follows:1.Designed a 2D model PGD-UNet based on two-dimensional convolution.PGD-UNet is mainly aimed at the problem of large changes in pancreas and tumor shapes and irregular boundaries.It integrates deformable convolution to model the spatial geometric transformation of the target.At the same time,this paper proposes an auxiliary positioning path to guide the offset learning of deformable convolution.The auxiliary positioning path provides a larger receptive field through hole convolution,and uses Coord Conv and Coord Pool to explicitly encode position information.Aiming at the label noise,this paper proposes the noise suppression loss function NSFL,which can effectively suppress the over-fitting phenomenon caused by noise while dealing with the problem of category imbalance.2.Designed a 3D model Dense UNet based on 3D convolution.Compared with the 2D model,the 3D model can extract the inter-layer dependencies of adjacent slices.Aiming at the problem of high memory requirements for 3D models,this paper adopts the idea of dense connection to design the feature extraction module of Dense UNet.Dense connection reduces the amount of model parameters through feature multiplexing,ensuring that the network has low redundancy,and it is not prone to overfitting when training on the pancreatic data set.In the densely connected feature extraction module,this paper uses a sawtooth wave method to stack the cavity convolution,so that the receptive field can be expanded quickly and effectively.3.Combining the 2D model and the 3D model,the cascading algorithm framework Dense 3D Cascade is designed.The cascading algorithm consists of one-stage 2D coarse segmentation and two-stage 3D fine segmentation.In the first stage,the approximate area of the target and the coarse segmentation result are obtained through PGD-UNet.In the second stage,three-dimensional image blocks are cut in the target area,and the coarse segmentation results are used as prior information,and they are sent to Dense UNet for fine segmentation.Finally,this paper also proposes a center-weighted strategy to fuse segmentation results to eliminate splicing traces.The cascade algorithm can effectively combine the advantages of the 2D model and the 3D model,and the segmentation performance is significantly improved.This article uses the published pancreas data set MSD-Pancreas to evaluate the proposed method.The experimental results show that: compared with the existing work,the method in this paper has achieved a significant improvement in the segmentation performance of pancreas and pancreatic tumors,and the segmentation results are more continuous and complete.