Tu Xiaoguang, Gao Jingjing, Zhu Chongjing, Cheng Jie-Zhi, Ma Zheng, Dai Xin, Xie Mei
School of Communication and Information Engineering, University of Electronic Science and Technology of China, Chengdu, 611731, Sichuan, China.
School of Electronic Engineering, University of Electronic Science and Technology of China, Chengdu, 611731, Sichuan, China.
Med Biol Eng Comput. 2016 Dec;54(12):1807-1818. doi: 10.1007/s11517-016-1540-7. Epub 2016 Jul 4.
Though numerous segmentation algorithms have been proposed to segment brain tissue from magnetic resonance (MR) images, few of them consider combining the tissue segmentation and bias field correction into a unified framework while simultaneously removing the noise. In this paper, we present a new unified MR image segmentation algorithm whereby tissue segmentation, bias correction and noise reduction are integrated within the same energy model. Our method is presented by a total variation term introduced to the coherent local intensity clustering criterion function. To solve the nonconvex problem with respect to membership functions, we add auxiliary variables in the energy function such as Chambolle's fast dual projection method can be used and the optimal segmentation and bias field estimation can be achieved simultaneously throughout the reciprocal iteration. Experimental results show that the proposed method has a salient advantage over the other three baseline methods on either tissue segmentation or bias correction, and the noise is significantly reduced via its applications on highly noise-corrupted images. Moreover, benefiting from the fast convergence of the proposed solution, our method is less time-consuming and robust to parameter setting.
尽管已经提出了许多分割算法来从磁共振（MR）图像中分割脑组织，但其中很少有算法考虑将组织分割和偏置场校正结合到一个统一的框架中，同时去除噪声。在本文中，我们提出了一种新的统一MR图像分割算法，该算法将组织分割、偏置校正和噪声降低集成在同一个能量模型中。我们的方法通过引入到相干局部强度聚类准则函数中的全变差项来表示。为了解决关于隶属函数的非凸问题，我们在能量函数中添加辅助变量，这样就可以使用Chambolle的快速对偶投影方法，并通过交替迭代同时实现最优分割和偏置场估计。实验结果表明，所提出的方法在组织分割或偏置校正方面相对于其他三种基线方法具有显著优势，并且通过将其应用于高噪声图像，噪声得到了显著降低。此外，受益于所提出解决方案的快速收敛，我们的方法耗时更少且对参数设置具有鲁棒性。
