基于FESS的混合模型脑图像分割方法  

Brain image segmentation method with hybrid model based on FESS

作　　者：连远锋[1] 赵剡[1] 何晖光[2] 陈雪姣[2] 

机构地区：[1]北京航空航天大学仪器科学与光电工程学院,北京100191 [2]中国科学院自动化研究所,北京100190

出　　处：《仪器仪表学报》2013年第6期1226-1232,共7页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(60701021);中国石油大学(北京)基金(KYJJ2012-05-21)资助项目

摘　　要：提出一种基于FESS的混合模型脑图像分割方法。其特点在于生成模型与判别模型得到了有效结合。生成模型通过条件随机场融合体素点的灰度信息、形状信息以及区域相邻关系,实现对脑子结构外观特征的描述。在此基础上,利用生成模型将训练样本映射到FESS特征空间;判别模型中采用LS-SVM分类器并将数据场应用于混合分割模型的训练过程中,降低了判别模型由于训练数据不平衡而引起的性能波动并提高其泛化能力。实验结果表明,与若干前沿的脑图像分割方法相比,该方法具有更好的分割质量和性能。This paper presents a new brain image segmentation method based on free energy score space （FESS） ,which combines the advantages of both generative model and discriminative model. Through conditional random field, the gen- erative model fuses the grey scale information, shape information and spatial neighborhood relation of the volume pixels ; and the appearance feature description of subcortical structure is realized. Based on the above,the generative model is used to map the training samples to FESS feature space. Furthermore;in the discriminative model, the LS-SVM classifier is used and the data field is applied in the training process of the mixed segmentation model, which reduces the per- formance fluctuation of the discriminative model induced by the imbalanced training data sets and effectively improves the generalization capability. Experimental results show that the proposed model possesses better segmentation quality and performance compared with several other state-of-the-art brain image se^nentation approaches.

关 键 词：FESS 生成模型 判别模型 图像分割 最小二乘支持向量机 数据场 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]