基于最优值搜索框架的CT序列图像肝脏分割

Liver Segmentation by Using an Optimal Framework for CT Images

显式模型下建立的肝脏分割算法受适用条件的限制,不能有效地控制分割进程,制约了算法鲁棒性和泛化能力的提高.针对这一问题,该文提出了一个新的方法,将肝脏分割问题转化为隐式函数最优值搜索,依据图像实际数据和先验形状信息有效地控制分割进程,以迭代搜索策略得到肝脏的最优分割结果.每一个迭代步骤由两部分组成:首先,利用MRF(Markov Random Field)实现已有肝脏轮廓邻域的局部再分割,重新调整肝脏轮廓;然后,利用先验形状稀疏表示调整后的形状,有效地修正噪声、边界模糊等因素引起的错误分割,并将修正后的肝脏轮廓用于下一轮局部再分割.肝脏分割结果在迭代过程中不断逼近真实值.为了提高形状稀疏表示的计算精度和速度,该文改进了相关技术细节,提出了动态字典生成策略,并利用混合粒子群算法求解稀疏表示方程.与金标准相比,该文所提算法在边界模糊、对比度低、存在大量其他干扰组织区域等不利因素的情景中,其分割精度(Volume Overlap Percentage,VOE)可达到90%以上.

Liver segmentation methods based on explicit model is restricted by many applying factors and can’t control the segmentation process in many cases,which limit the algorithm performance of the robustness and generalization.To this end,a novel method is proposed in this work,where the liver segmentation is transformed into the optimization of implicit function.In our work,the segmentation process is managed by image data and shape priors,and the optimal liver segmentation result is obtained through iteration steps.There are two parts for every iteration.Firstly,a new liver contour is gained by re-segmenting the liver within the neighbor of the contour points based on MRF.And then,for correcting the segmentation errors caused by edge fuzzy and noise,et al.,the new liver shape which is obtained in the previous step is adjusted by sparse shape representation based on shape priors.The revised liver contour is transferred into next iteration step.The liver segmentation get closed to the ground truth during the iteration process.Moreover,in order to improve the accuracy and speed of the sparse shape representation process,dynamic shape dictionary and HPSO （Hybrid Particle Swarm Optimization）are used fornbsp;sparse shape representation solution.Comparing with the Ground Truth,the segmentation accuracy （VOE,Volume Overlap Percentage）reaches as high as 90% under the interference of fuzzy edge,low-contrast,noisy,and other adverse factors.