基于SAE深度特征学习的数字人脑切片图像分割

发布时间：2018-05-11 04:21

本文选题：中国可视化人体数据集 + 脑组织分割　；参考：《计算机辅助设计与图形学学报》2016年08期

【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.
[Abstract]:Aiming at the shortcomings of the present segmentation algorithm based on digital human brain slice image, such as less segmentation accuracy and lower efficiency, a new segmentation algorithm based on sparse self-encoder (SAE) depth feature learning is proposed. In the stage of feature extraction, SAE is trained from coarse to fine to enhance the discriminative ability of the depth features learned by the model. In the stage of classification, the softmax classifier is used to segment the target. The experimental results of white matter segmentation and 3D reconstruction of Chinese visualized human body CVH data set show that compared with other traditional manual features, such as image intensity features, The feature of directional gradient histogram and the feature of principal component analysis (PCA) extracted by SAE have stronger discriminant ability and higher segmentation accuracy.
【作者单位】：重庆大学光电技术及系统教育部重点实验室;重庆大学光电工程学院;重庆师范大学计算机与信息科学学院;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;
【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)
【分类号】：TP391.41

【相似文献】