U形网络的脊椎分割改进方法研究  

Improved spinal segmentation method based on U⁃net

作　　者：梁淑芬[1] 杨芳臣 秦传波[1] LIANG Shufen;YANG Fangchen;QIN Chuanbo(Wuyi University,Jiangmen 529020,China)

机构地区：[1]五邑大学,广东江门529020

出　　处：《现代电子技术》2020年第21期31-34,38,共5页Modern Electronics Technique

基　　金：国家自然科学基金(33518001);国家自然科学基金(61372193);广东省自然科学基金(S2013010013311);广东省高校省级重点平台和重大科研项目特色创新项目(2015GXJK151,2015KTSCX148);自动化特色专业(GDJX2017002);江门市科技计划项目(江科[2017]268);五邑大学青年科研基金(2015zk10);五邑大学校级教学质量与教学改革工程项目(JX2018009)。

摘　　要：针对当前U-Net网络模型以及相关卷积网络在脊椎分割中分割边缘精度低,对目标的识别率低等问题,提出几种改进的语义分割网络模型。考虑到脊椎MRI图像中目标区域连续且集中,并存在复数微小区域,边缘特征丰富,运用新型卷积块替换标准卷积结构,利用多路径思想,融入复数编解码器结构,降低运算复杂度,提升边缘特征提取力度,应对不同的特征提取问题。模型在SpineWeb的High anisotropy MRIs of the lower back数据集2000张脊椎图像上做实验,运用训练好的模型对同组脊椎MRI图像作预测。实验结果显示,网络模型的预测结果与提供的真值标签在Dice系数以及Precision系数等几项评价指标上分别达到0.891和0.894。As the current U⁃net models and related convolution networks have low accuracy of edge segmentation in spinal segmentation,several improved semantic segmentation network models are proposed.Since the continuity and concentration of target areas in the spine MRI(magnetic resonance imaging)images,plural small areas and rich edge features are considered,a new convolution block is used in the method to replace standard convolution structure.Several codec structures are integrated with an idea of multi⁃path,which can reduce the computation complexity and improve segmentation edge feature extraction to deal with different feature extraction.The experiments of 2000 spinal images in SpineWeb′s High Anisotropy MRIs of the Lower Back dataset were carried out,using the trained model to forecast the spine MRI images in the same group.The experimental results show that the prediction results of the network model and evaluation indexes on the provided truth⁃value label for Dice coefficient and Precision coefficient have reached 0.891 and 0.894.

关 键 词：脊椎分割 网络模型 U形网络 结构替换 特征提取 图像预测 

分 类 号：TN911.73-34[电子电信—通信与信息系统]