Research On Multi-Modal MRI Image Segmentation Based On Deep Learning

Glial brain tumors are the most common primary malignant tumors.Once a patient is diagnosed with brain tumors,their life will be seriously threatened.Magnetic Resonance Imaging(MRI)technology is of great significance for the clinical diagnosis and medical research of brain tumors and other diseases.Manual segmentation of brain tumors by experts is a professional,time-consuming task and highly dependent on the personal experience of the experts.Therefore,the use of computer algorithms to assist brain tumor segmentation plays an important role in computer-aided diagnosis.With the development of deep learning,various deep neural network models have been proposed,and the emergence of these networks provides new ideas for brain tumor segmentation on MRI.This paper will design an efficient and accurate neural network model for brain tumor segmentation tasks from two dimensions and three dimensions.The specific work is as follows:Firstly,a PAU-Net model,a multimodal brain tumour segmentation network based on 2D convolution,is proposed and a multi-view fusion algorithm based on a 2D segmentation network is designed.The PAU-Net model not only inherits the ideas of the U-Net network structure and residual module,but also adds a pyramidal residual convolution module,which enhances the perceptual field and information extraction ability of the model and improves the model’s ability to discriminate information The multi-view fusion algorithm combines the advantages of segmentation capability of sliced images of different orientations on the same network and further improves the segmentation accuracy from the perspective of data dimensionality.Secondly,we propose a 3D U-Net-based multimodal brain tumour segmentation network,PTCBAMU-Net,to increase the brain tumour segmentation method from 2D to 3D scales,and achieve end-to-end segmentation of 3D multimodal brain tumour medical images.The network improves the pyramidal residual structure proposed in the 2D segmentation network into a 3D convolutional structure and introduces a new global attention module and a dual attention module to further improve the segmentation accuracy.Experiments show that the method proposed in this paper can effectively improve the segmentation of brain tumours and is competitive in 3D multimodal brain tumour MRI image segmentation methods.Finally,the above multimodal brain tumour segmentation algorithm was deployed and tested on a Jetson TX2 embedded motherboard to design a multimodal MRI brain tumour segmentation system,which can effectively improve the efficiency of clinicians’ diagnosis by achieving multi-viewing and accurate segmentation functions for multimodal brain tumour images.