基于Ghost Module的轻量级脑肿瘤3D MRI分割研究  

Lightweight 3D MRI segmentation study of brain tumors based on Ghost Module

作　　者：刘丽伟[1] 赵强 LIU Liwei;ZHAO Qiang(School of Computer Science&Engineering,Changchun University of Technology,Changchun 130102,China)

机构地区：[1]长春工业大学计算机科学与工程学院,吉林长春130102

出　　处：《长春工业大学学报》2022年第6期686-692,共7页Journal of Changchun University of Technology

基　　金：吉林省教育厅科学技术研究规划项目(JJKH20210738KJ)。

摘　　要：针对当下三维CNN对脑肿瘤MRI分割时训练参数量过大、网络收敛性差、训练时间过长的问题,提出S-GG Net。先通过对轻量级结构Ghost Module进行改进形成Shuffle-Ghost Module;并以此为基础单元形成两种残差结构,网络仿造U-Net结构分别对网络的压缩路径、扩展路径进行重新构建。在公开的脑肿瘤数据集BraTs进行实验得知,文中方法在大幅缩减参数量的情况下,仍然能够保证训练效果。To address the current problems of excessive amount of training parameters,poor network convergence and long training time in 3D CNN for brain tumor MRI segmentation,we propose S-GG Net.firstly,the Shuffle-Ghost Module is formed by improving the lightweight structure Ghost Module;and based on this,the network is reconstructed to form two residual structures.The unit forms two residual structures,and the network imitates the U-Net structure to reconstruct the compressed path and extended path of the network respectively.Experiments on BraTs,a publicly available brain tumor dataset,show that this method can guarantee the training effect even when the number of parameters is reduced.

关 键 词：脑肿瘤分割 3D MRI 轻量级改进 S-GGNet 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]