Research On Segmentation Method Of Liver CT Image Based On Improved U-Net

Liver cancer is one of the most prevalent cancers worldwide,which seriously endangers people’s life and health.Early screening and treatment can effectively reduce the incidence and mortality of cancer.Currently,the most commonly used method for examining liver tumors is Computed Tomography(CT).The main methods for treating tumors include tumor resection,interventional therapy,and radiation therapy.Obtaining detailed information on the number and size of tumors before surgery is of great significance for the scientific formulation of surgical plans.Therefore,ensuring accurate segmentation is crucial before conducting liver tumor treatment.At present,there are multiple difficulties in the segmentation of liver tumors: there are many organs and blood vessels distributed in the vicinity of the liver;the location,size and shape of liver tumors are very complex,with unclear boundaries and uneven grayscale.In medical diagnosis,manual segmentation requires a significant investment of time and effort,and the results may be inconsistent.The semi-automatic segmentation method requires manual intervention,and its effectiveness is influenced by personal experience and subjectivity.Currently,deep learning technology has been widely used in the field of medical image segmentation.This article is based on the U-Net network framework for research.Aiming at the problems of insufficient information extraction and low segmentation accuracy in traditional U-Net network,a method of liver segmentation based on improved U-Net is proposed.Firstly,in the U-Net encoding stage,the hybrid dilated convolution is used to replace the original convolution block to expand the receptive field and extract more context information;In the decoding stage,dense upsampling convolution is used to capture and decode more detailed information;The residual module is introduced to speed up the training of the model and prevent network degradation.Secondly,the attention mechanism(CBAM)is added between each jump connection to focus the model on the feature information of the region of interest,and combined with Focal Tversky loss function to improve the class imbalance problem.Experiments on the Li TS2017 dataset show that compared with traditional U-Net,the Dice index of the proposed method in liver segmentation has been improved by 3.56%,the recall rate has been improved by 3.71%,and the accuracy rate has been improved by 2.76%.Aiming at feature learning of small and medium-sized liver tumor regions in CT images,a method of liver tumor segmentation based on improved UNet3+ is proposed: the receptive field block is introduced in the encoding phase of the UNet3+ to enhance the feature extraction ability;in the decoding stage,data-dependent upsampling(DUpsampling)is used to replace the original upsampling mode to restore more detailed image feature information;the normalized attention mechanism(NAM)is added between jump connections to focus the model on the features of the target region and suppresses redundant features;finally,the hybrid loss function combining cross entropy and Dice is used to improve the class imbalance problem and improve the convergence performance of the network.Experiments on the Li TS2017 dataset show that compared with original UNet3+,the Dice index of the proposed method in liver tumor segmentation has been improved by 4.05%,the recall rate has been improved by 2.83%,and the accuracy rate has been improved by 3.92%.