Research On The Segmentation Method Of Pancreas And Pancreatic Lesions Based On Convolutional Neural Network

Currently,doctors often use computed tomography(CT)imaging to diagnose pancreatic cancer.But this is very dependent on the doctor’s personal experience,timeconsuming and labor-intensive.Assisting doctors to make accurate judgments on CT images is a problem that needs to be solved urgently.Thanks to the development of computer computing power and the rise of image processing technology,computer-aided diagnosis technology based on deep learning has gradually been applied to clinical medicine.In order to effectively assist doctors in the detection of pancreatic cancer,this thesis uses convolutional neural networks to segment the pancreas and extract the lesion area.The main research contents of this thesis are as follows:1.To solve the problem of inconsistency within pancreas and no difference between categories,this thesis proposes a multi-scale dense residual attention U-shaped network segmentation algorithm that strengthens context information and semantic information extraction.This method introduces multi-scale convolution at the low level,and combines the information flow of different receptive fields to obtain more shape information.At the same time,the binary cross entropy(BCE)loss function is introduced to strengthen the extraction of edge information.Effectively alleviate the problem of no difference within the category.At each level of the U-shaped network,dense connections are used,and a channel attention mechanism is introduced between skip connections to strengthen the learning of semantic information,and effectively solve the problem of inconsistency in the class.This thesis conducted four cross-validation experiments on National Institutes of Health(NIH)dataset and MICCAI Decathlon Challenge(MSD)dataset.Experimental results show that on the NIH and MSD datasets,the average Dice similarity coefficients of pancreas segmentation are 86.10%±3.52% and 88.50%±3.70%,respectively.2.To solve the extremely unbalanced problem in the segmentation of pancreas and lesions,this thesis proposes a multi-scale dual-branch weight sharing fusion strategy.The first branch is mainly used to extract normal pancreas.In the low-level stage of coding,multi-scale is used to obtain shape information.In the decoding stage,bilinear interpolation combined with convolution is used to replace deconvolution,which effectively guarantees the enrichment of semantic information in the upsampling process.In the second branch,U-Net is used to extract pancreatic cancer information,and feature channel fusion is performed with the first branch at each skip connection stage to perform weight sharing.At the same time,this thesis constructs a loss function that assigns learning weights by category.In the process of backpropagation,the sum of the category pixels is reversed,so that the weight of the pancreatic cancerous part is larger,which is helpful to solve the problem of category imbalance.This thesis conducted four crossvalidation experiments on the MSD dataset.The test results showed that the average dice similarity coefficients of pancreas and cancerous regions were 83.92%±3.13% and 52.32%±2.88%,respectively.