面向3D CT影像处理的无监督推荐标注算法  

An Unsupervised Suggestive Annotation Algorithm for 3D CT Image Processing

作　　者：冯浩哲 张鹏[3] 徐欣楠 郝鹏翼[4,2] 吴福理 吴健[5,2] 陈为[1,2] Feng Haozhe;Zhang Peng;Xu Xinnan;Hao Pengyi;Wu Fuli;Wu Jian;Chen Wei(State Key Laboratory of CAD&CG,Zhejiang University,Hangzhou 310058;School of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310014;School of Computer Science and Technology,Zhejiang University,Hangzhou 310007;RealDoctor Artificial Intelligence Research Center,Zhejiang University,Hangzhou 310007;Department of Thoracic Surgery,Shanghai Pulmonary Hospital,Tongji University,Shanghai 200433)

机构地区：[1]浙江大学CAD&CG国家重点实验室,杭州310058 [2]浙江大学睿医人工智能研究中心,杭州310007 [3]同济大学附属上海市肺科医院胸外科,上海200433 [4]浙江工业大学计算机科学与技术学院,杭州310014 [5]浙江大学计算机科学与技术学院,杭州310007

出　　处：《计算机辅助设计与图形学学报》2019年第2期183-189,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61772456;61801428;61672453;81470244;91642108;81622001);浙江省自然科学基金(LY18F020034);中国教育部科技司科技创新平台培育项目(2017PT18);浙江大学教育基金(K18-51120-004;K17-51120-017;K17-518051-021);之江实验室重大科研项目(2018DG0ZX01);上海市申康发展中心三年行动计划(SHDC1201606);上海市卫计委新百人(2017BR026);上海市人才发展基金(17SG23)

摘　　要：在3D CT影像分析上应用深度学习技术时,通常需要采用交互标注工具标注一组训练数据.针对3D CT影像一般包含数量较多的切片,医学影像交互标注工作量非常巨大且标注成本非常高的问题,提出一种面向3DCT影像数据交互标注的无监督推荐标注算法,通过构造稠密深度自动编码器DCDAE (densely-connected deep auto encoder)提取3D影像的高层特征,同时采用密度-谱聚类来筛选最具标注价值的影像,从而极大减少需要标注的数据量.算法提出了全自动的推荐标注流程,在提取图像特征时采用稠密连接结构改进DCDAE,减少了参数量并使得提取的特征更有区分度,同时对特征采用密度-谱聚类算法进行孤立点鉴别,并依据相关性矩阵自适应调整聚类个数;在肺结节语义分割任务上采用LIDC-IDRI数据集对算法进行了实验.It is necessary to interactively annotate a set of training data when applying deep learning technology on 3D CT images.This requires a huge amount of workload by medical experts for manual annotations.This paper proposes an unsupervised suggestive annotation algorithm for 3D CT images that employs two new techniques(densely connected deep auto encoder and density-spectral clustering)to significantly reduce annotation requirements.Our algorithm results in three advantages.First,it is fully unsupervised.Second,a new auto-encoder named DCDAE is proposed to reduce the amount of model parameters and extract discriminative features by combining deep autoencoder and dense connection structure.Third,a new clustering algorithm named density-spectral clustering is proposed to find the outliers and automatically adjust the cluster number according to the affinity matrix of the dataset.The algorithm is applied on lung nodule semantic segmentation task using LIDC-IDRI dataset.

关 键 词：医学影像分析 深度自动编码器 聚类算法 推荐标注算法 交互标注 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]