一致性正则化与代理标签的骨骼点云半监督分割  

Consistent Regularization and Proxy Label based Bone Point Cloud Semi-Supervised Segmentation

作　　者：周长虹 蒋俊锋 张文玺[2] 黄瑞 张昊[3] 陈亮 ZHOU Changhong;JIANG Junfeng;ZHANG Wenxi;HUANG Rui;ZHANG Hao;CHEN Liang(College of Internet of Things Engineering,Hohai University,Changzhou 213022;Department of Orthopedics,Liyang People's Hospital,Changzhou 213300;Department of Vascular Surgery,Changhai Hospital,Shanghai 200433;Changzhou Jinse Medical Information Technology Co.,Ltd.,Changzhou 213000)

机构地区：[1]河海大学物联网工程学院,常州213022 [2]溧阳市人民医院骨科,常州213300 [3]上海长海医院血管外科,上海200433 [4]常州锦瑟医疗信息科技有限公司,常州213000

出　　处：《计算机与数字工程》2022年第10期2223-2228,共6页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:52075148)资助。

摘　　要：计算机辅助骨科术前规划中,需要提取碎骨外表面并对骨骼外表面划分语义区域。基于监督学习的点云分割方法手工标注成本较高,费时费力。结合一致性正则化与代理标签的半监督分割方法两者的优点,提出一种针对人体骨骼语义分割的新型半监督学习方法。该方法从学生网络的输出中过滤出置性度较高的结果作为无标签数据的代理标签,用于和教师网络的输出做比对,从而训练学生网络的参数,然后根据学生网络的参数,更新教师网络的参数。实验表明,利用10%标注数据即可接近100%标注数据的精度,碎骨内外表面分割的平均交并比达到84.2%,股骨表面语义区域分割平均交并比达到94.3%。该方法能有效减少数据标注,从而降低数据标注成本并提高传统半监督学习方法的精度。In computer-aided orthopedic pre-operative planning,the outer surface of the bone piece needs to be extracted and the bone surface needs to be segmented. Point cloud segmentation methods,based on supervised learning,are costly and time-consuming. A novel semi-supervised method is proposed to semantically segment the point cloud by incorporating both consistent regularization and proxy label. This method keeps the results with higher probability from the output of the student network as proxy labels for the unlabeled data. Then,the selected proxy labels are used to compare with the output of the teacher network.Thus,the student network is trained. Next,the parameters of the teacher network are updated concurrently with respect to the student network. The experiments demonstrate that using 10% labeled data can be comparable to 100% labeled data in accuracy. The extraction of bone piece outer surface reaches 84.2%,and the femur surface segmentation reaches 94.3%. The method not only effectively reduces data annotation and cost,but also improves the accuracy of traditional semi-supervised learning methods.

关 键 词：半监督学习 骨骼语义分割 点云 一致性正则化 代理标签 

分 类 号：R687.3[医药卫生—骨科学]