Research On Liver And Tumor CT Image Segmentation Method Based On U-Net Mode

Liver cancer is a common cause of death in humans.Currently,the effective method for treating liver cancer is to surgically remove the liver tumor in the early stages of liver cancer.Liver tumors are diagnosed clinically mainly by CT images.Therefore,accurately segmenting the liver and liver tumor regions in CT images is crucial for timely detection of liver cancer and providing positional information reference for subsequent surgeries.In order to achieve this goal,people have been studying the application of computer image processing technology to the segmentation task of liver and liver tumors.However,in CT images,the liver and liver tumors have the characteristics of variable shape,size,and blurred boundaries,especially the size difference of liver tumors is very large,and the number of liver tumors is also random.Based on these difficulties,the most commonly used U-Net model in the field of medical image segmentation still has shortcomings such as over segmentation,under segmentation,inaccurate boundary segmentation,and inability to recognize and segment some small tumors in the segmentation task of liver and liver tumors.In order to solve the bottleneck of the U-Net model,this article redesigns the segmentation model based on U-Net,and the specific work is as follows:(1)Based on U-Net,a segmentation model(RISU-Net)is designed,which integrates residual path module and channel attention module and has the ability of multi-scale feature extraction.This segmentation model combines multi-scale feature extraction module Inception with channel attention SE module and replaces the ordinary convolution of U-Net,so that the model can better focus on important features while having multi-scale feature extraction capability;The introduction of the residual path module at the U-Net jump connection mitigates the incompatibility between the encoder and decoder features and makes better use of the extracted features.On the Li TS2017 data set,it is verified that RISU-Net has some improvement over U-Net in over-segmentation and under-segmentation of liver and liver tumors.(2)Based on U-Net,a multi-scale feature segmentation model(PCRU-Net)with pyramid residual convolution,channel and spatial dual attention is designed.The model combines pyramid residual convolution with CBAM attention on the basis of preserving the residual path of RISU-Net,and replaces the ordinary convolution of UNet,so that the model has stronger multi-scale feature extraction ability than RISU-Net and can better focus on important features in both channel and space dimensions;The loss function is redesigned,and the cross entropy loss function is combined with the Dice loss function to improve the imbalance between the number of segmentation objects and background categories.On the Li TS2017 data set,it is verified that PCRUNet has further improved the segmentation accuracy compared with RISU-Net,and has improved the problem that the tumors at the liver boundary cannot be segmented normally due to the insufficient accuracy of boundary segmentation,and the smaller tumors cannot be segmented normally.