基于SAE深度特征学习的数字人脑切片图像分割  

Deep SAE Feature Learning Based Segmentation for Digital Human Brain Image

作　　者：赵广军[1,2] 王旭初[1,2] 牛彦敏[3] 谭立文 张绍祥[4] 

机构地区：[1]重庆大学光电技术及系统教育部重点实验室,重庆400040 [2]重庆大学光电工程学院,重庆400040 [3]重庆师范大学计算机与信息科学学院,重庆401331 [4]中国人民解放军第三军医大学生物医学工程学院数字医学研究所,重庆400038

出　　处：《计算机辅助设计与图形学学报》2016年第8期1297-1305,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(60903142;61190122);中国博士后基金特别资助(2013T60841);中央高校基本业务费项目(106112015CDJXY120003)

摘　　要：针对目前基于数字人脑切片图像的分割算法较少,分割精度和有效性较低等不足,提出一种基于稀疏自编码器(SAE)深度特征学习的分割算法.在特征提取阶段,采用从粗到精两级方式对SAE进行训练,以增强模型学习到的深度特征的鉴别能力;在分类阶段,使用softmax分类器进行目标分割.对中国可视化人体(CVH)数据集的脑白质分割及三维重建的实验结果表明,相对于其他传统的手工特征(如图像强度特征、方向梯度直方图特征和主成分分析特征),SAE提取的图像深度特征具有更强的鉴别能力,显著地提高了分割精度.There are few algorithms for segmenting cryosection brain images, and most existing segmentation techniques presented limited precision and low efficiency. To address these problems, this paper proposed a novel deep feature learning-based segmentation algorithm using sparse autoencoder（SAE）. At the stage of feature extraction, SAE is trained twice to enhance the discriminability of the deep-learned feature representations. At the stage of classification, a softmax classifier is used for segmenting different objects. Experimental results of white matter segmentation on the Chinese Visible Human（CVH） dataset and its 3-D reconstruction show that, the learned deep feature performs much better in discriminability compared with other representative hand-crafted features（such as intensity, histogram of oriented gradient and principal components analysis） and achieves higher recognition accuracy.

关 键 词：中国可视化人体数据集 脑组织分割 稀疏自编码器 深度特征 softmax分类器 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]