Research On MRI Liver And Liver Tumor Segmentation Model Based On Deep Learning

In recent years,deep learning method has made a lot of rapid development in the field of image segmentation,and it has been proved that it has better effect than traditional image segmentation methods in many experiments.In addition,it has the characteristics of end-to-end prediction,no need to introduce a large number of hyperparameters,and automatic feature extraction,and has been applied to many practical scenarios.Medical image processing is a popular research direction in computer vision.Image segmentation based on medical images is of great significance for actual clinical diagnosis.Segmentation of medical images using deep learning can provide doctors with key information of patients’ lesion sites quickly and accurately,and provide some help for doctors in disease diagnosis and clinical treatment.Due to the scarcity of public data sets on liver and tumor in MR medical images,we cooperated with a hospital in Hunan Province to organize and prepare to publish an MR liver tumor dataset to fill the gap in the MR field of liver tumor datasets.The data set collected abdominal T2-weighted MR images of 103 patients with hepatocellular carcinoma(HCC).Each data contains liver and liver tumors,and their labels are annotated by a professional radiologist.Considering some difficulties of deep learning method in tumor segmentation,such as the difficulty of segmentation due to the small size of liver tumors,insufficient use of liver information in the previous networks,a huge amount of computation of neural network and insufficient local feature extraction,we propose a new tumor segmentation network and a new loss function based on the new data set.The main contents of this paper are as follows:(1)In order to solve the problems that a single size convolution kernel cannot obtain a multi-scale receptive field and the 3D neural network requires a large amount of computation,we choose a lightweight network as the basic network of this paper,the basic network used combines multi-scale receptive field and dilated convolution,so that the network can extract the receptive field features of different scales and obtain more information.The network also uses grouped convolutions to drastically reduce the amount of network parameters.We improved and added an attention module to the basic network to improve the learning strength of the network for effective features,and finally make the model achieve better performance.(1)In order to solve the problems that a single-size convolution kernel cannot obtain a multi-scale receptive field and the 3D neural network requires a large amount of computation,we improve an existing network and propose a new network which uses a combination of multi-scale receptive fields and dilated convolution to simultaneously extract the features of receptive fields at different scales for more information.In addition,we also added an attention mechanism to the new network,which makes the network pay more attention to the learning of segmentation targets through the weighted summation of features.The group convolution used in the network greatly reduces the network parameters,and finally a lightweight network is obtained.(2)Aiming at the difficulty of small target segmentation,we propose a new multiscale local loss function based on cross-entropy loss.The loss function emphasizes the local features around the segmentation target learned by the network,which is beneficial for the network to find and locate small targets.Perform segmentation,and experiments show that the combination of multi-scale local loss and original global loss improves the segmentation accuracy of each network.(3)Considering issues such as patient privacy and data labeling,medical imaging datasets are mostly small sample datasets.However,the previous methods have the problem of insufficient utilization of liver information.when segmenting liver tumors,as a result,the network cannot learn the data features better.We propose a two-stage segmentation network to segment tumors.The network introduces liver information into the network through liver feature filtering,allowing the model to learn more liver internal features for better tumor segmentation.Experiments show that our proposed two-stage segmentation network shows better performance on tumor segmentation.