Research On Liver Tumor Image Segmentation Based On MSM

With the rapid development of magnetic resonance imaging,computed tomography and other medical imaging technologies,medical images,as the key basis for diagnosis and treatment,are playing an increasingly important role in clinical medicine.Liver cancer is a malignant disease with high fatality rate.It is very important to use computed tomography for medical diagnosis and treatment.With the rapid development of deep learning,automatic segmentation of liver lesions through convolutional neural network can assist medical workers in better diagnosis and treatment.However,in many cases,the fault of the hardware equipment may lead to the problems such as interference information and fuzzy tumor boundary.This not only directly affects the manual diagnosis and treatment,but also affects the accuracy of neural network segmentation.Due to the poor quality of some medical images,this paper proposes a liver tumor segmentation method based on convolutional neural network,which takes the super-resolution of medical images of liver tumors as the auxiliary task.The MSMN method proposed in this paper realizes the super-resolution.On the one hand,it can provide medical workers with better visual quality liver tumor images.On the other hand,it provides multi-scale resolution image input for the MSCMT segmentation algorithm proposed in this paper.MSMN is a super resolution algorithm based on multi-scale feature extraction and multi-branch training.MSMN is composed of a main path composed of 2D module DCN,and two branches composed of 3D modules TCN1 and TCN2 with different convolution kernel sizes.The 2D trunk path extracts shallow and low frequency information,while the 3D branch path extracts more abstract and high frequency information.In the forward propagation process,the information of the two branches is further pooled in the 2D trunk through the multi-scale pooling module MSP.This strategy of multi-branch training and local information fusion improves the ability of information extraction and fusion.In addition,the multi-scale pooling module MSP designed in this paper further strengthens the feature extraction capability of different scales.Through experiments on liver tumor data sets,MSMN is proved to have better performance than the other four advanced methods from both quantitative and qualitative aspects.MSCMT is an image segmentation algorithm based on multi-scale information input,convolutional neural network and transformer.The core idea of MSCMT is to make use of prior liver tumor image information extracted by MSMN,and use super resolution reconstructed images of different scales as network input.Through feature extraction of multi-source information and images of different scales,feature fusion is carried out,and finally realize effective segmentation of liver tumor images with poor quality and low resolution.In addition,the method of extracting local information by convolutional neural network and extracting global information by Swin-T is combined to expand the receptive field of the network.A wealth of experiments are carried out in this paper.By comparing with the other five methods,it is verified that the MSCMT network has superior performance in the field of segmentation,in addition,it has better robustness and universality for images with poor input quality.In a word,the liver tumor segmentation method proposed in this paper,which takes the liver tumor medical image super resolution as the auxiliary task,not only improves the visual quality of the image,but also shows superior performance and great application prospect.