基于改进U-Net模型下的椎骨CT图像分割方法
2021-09-06刘侠余鸿波李冰王波
刘侠 余鸿波 李冰 王波
摘 要:针对经典U-Net模型在椎骨CT分割过程中对图像信息利用不充分而造成图像边缘分割不清的问题,采用一种基于改进的U-Net模型的椎骨CT图像分割算法。首先在经典U-Net模型基础上进行了改进,其次利用改进的U-Net模型分割出椎骨区域并得到粗分割结果,最后对粗分割结果利用图割算法(Graph-Cut)加强边缘约束,从而做到边缘细化分割。分割方法整体分割精度可达到95.5%,Dice系数96.2%,Jaccard系数92.6%,HdD指标4.88%。与经典U-Net模型相比Dice系数提高2.2%,Jaccard系数提高3.7%,HdD指标降低13.9%。实验结果表明,提出的分割方法对椎骨图像分割可以达到精确分割的效果,能够适用于临床中的椎骨图像分割任务。
关键词:
椎骨CT图像;U-Net网络;Graph-Cut算法;图像分割
DOI:10.15938/j.jhust.2021.03.009
中图分类号: TP391.4
文献标志码: A
文章编号: 1007-2683(2021)03-0058-07
Vertebra CT Image Segmentation Method Based
on Improved U-Net Model
LIU Xia, YU Hong-bo, LI Bing, WANG Bo
(School of Automation, Harbin University of Science and Technology, Harbin 150080, China)
Abstract:In view of the problem that the classical u-net model does not make full use of the image information during the CT segmentation of vertebrae, which leads to unclear image edge segmentation, an improved algorithm of CT image segmentation of vertebrae based on the u-net model is proposed. Firstly, it was improved and optimized on the basis of the classic u-net model. Then, the improved u-net model was used to segment the vertebral region and obtain the rough segmentation result. Finally, the rough segmentation result was strengthened by the edge constraint algorithm (Graph-Cut) so as to achieve the edge refinement segmentation. In this paper, the segmentation precision of the methord can be reached 95.5%. The Dice coefficient is 96.2%. The Jaccard coefficient is 92.6%.The HdD index is 4.88. Compared with the classic u-net model, the Dice coefficient increased by 2.2%, the Jaccard coefficient increased by 3.7%, and the HdD index decreased by 13.9%.
Keywords:CT image of vertebrae; U-Net network; Graph-Cut algorithm; image segmentation
0 引 言
健康的椎骨维持着人体正常生理形态,从颅骨开始中间连接肋骨到髋骨为止,是人体后侧支撑,具有挺拔躯体,保护内脏器官等重要作用。然而現代社会中,由于人们生活节奏加快、工作压力繁重、从事人力劳动人口众多等原因,患有椎骨疾病[1-2]的患者不断呈现出年轻化的趋势并且病患数量也再递增,据统计在40岁以下人群中患有椎骨相关疾病的人数接近40%以上。还原患者椎骨真实形态可以辅助骨科医师降低手术过程中的风险。所以针对于椎骨CT图像高精度分割出有效的椎骨信息对于辅助临床诊断迫在眉睫。
为了解决椎骨分割精度问题,国内外学者已经提出许多方法用于椎骨的自动分割。Yao等[3]提出使用一个简单的阈值来获得初始的脊柱分割,然后采用分水岭算法和有向图搜索相结合的方法提取椎管,最后用四部分椎体模型(椎体、棘突、左右横突)将椎体区域分段,并与相邻的肋骨等结构分开。……
