面向3D CT影像处理的无监督推荐标注算法

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.

An Unsupervised Suggestive Annotation Algorithm for 3D CT Image Processing

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.