Research On Segmentation Method Of Brain Tumor Images Based On 3D Convolutional Neural Network

Brain tumors have a significant impact on patient’s lives and are a disease with a high mortality rate.Among them,glioma is the most common primary malignant tumor of the brain with a very low cure rate.The shape,location,and malignancy of brain tumors do not have universal general rules for different patients,so it is difficult to segment them.The brain tumor segmentation method relies on brain MRI technology,which needs to mark different tumor areas from normal tissues.Currently,manual segmentation is used in clinical diagnosis,which is not only time-consuming and labor-intensive but also the subjective experience and fatigue of physicians have an impact on the segmentation effect.To assist physicians to provide accurate and rapid diagnosis,this paper investigates the MRI segmentation method of brain tumors based on a 3D convolutional neural network,and the main research contents include:(1)Brain tumors do not have universal regularity in shape,location,and size,and cannot be enlarged for resection,so the acquisition of tumor boundary information is especially critical,and therefore the perceptual field during network feature extraction needs to be increased.In this context,a feature multiplexing segmentation algorithm with enhanced information extraction is proposed.Among them,the information feature extraction module is proposed: the perceptual field is increased by using null convolution to increase the extraction of information,and the parallel structure is used to compensate for the information loss caused by the tessellation effect of null convolution in combination with deep convolution.The shallow information is size-matched for feature activation and added with the deep information to achieve feature reuse.The experimental results show that the method improves Dice by 2.23% on average and PPV by 4.18% on average in the three segmentation regions compared with the segmentation results of U-Net in each tumor region.(2)Network methods with higher accuracy are often accompanied by a huge number of parameters and computation,and there is no advantage in performing specific model deployment,so lightweight research on segmentation methods is needed.Based on the V-Net structure of the high-precision segmentation network,a lightweight segmentation method based on the improved Ghost Module is proposed.This method proposes two lightweight modules for network construction by improving the lightweight module Ghost Module and combining the ideas of global attention mechanism and residuals to finally form the lightweight network S-GG Net.This network solves the problem of excessive training time caused by the excessive amount of training parameters for 3D MRI segmentation of brain tumors.Compared with the V-Net method,the S-GG Net method proposed in this paper is 1/6.7 of the latter in terms of the number of parameters and 1/8.0of the latter in terms of computation with similar accuracy,which achieves the original purpose of assisting doctors in accurate and rapid diagnosis.