基于Metropolis-SA算法的脑部磁共振血管造影图像分割

Cerebrovascular segmentation from magnetic resonance angiography based on Metropolis-SA algorithm

目的利用三维Markov随机场(MRF)模型分割脑部磁共振血管造影(MRA)。方法 MRF的似然概率采用了瑞利分布和高斯混合分布函数,并利用最大期望(EM)算法精确估计出混合参数;先验概率采用Ising-MRF模型,并利用误差试探法估计出正则化参数。为避免利用迭代条件模式(ICM)进行图像分割时常陷入局部最优解,实验提出了基于Metropolis采样算法的模拟退火(SA)技术。结果实现了三维MRF的全局最优解,分割模型可分辨3个体素的细小血管。临床数据采用南方医院影像中心提供的患者TOF-MRA数据(1.5TGEMRIscanner),空间分辨率0.43mm×0.43mm×0.50mm;原始数据的像素空间大小为512×512×128;实际采用的空间大小和分辨率分别为256×256×64和0.80mm×0.80mm×1.20mm。实验对每一套临床数据采用SA、ICM、MSA算法分别进行分割比较,分割结果存在有限差异,采用15步迭代计算的时间消耗分别为1 029 s、463 s、560 s。结论实验通过三维仿真数据分割结果表明,Metropolis-SA迭代求解算法能够实现更低的全局误差,并且实际脑部MRA数据的分割与最大密度投影相比较,反映出较好效果。

Objective To study segmentation of brain magnetic resonance angiography by three-dimensional Markov random field （MRF） model. Methods Rayleigh and Gaussian mixture distributions were adopted to calculate the likelihood probability and the mixture parameters were accurately estimated by expectation maximization （EM） algorithm. Ising-MRF model was applied for the calculation of prior probability and the regularization parameter was estimated using trial-and-error method. To avoid the occurrence of local optimum solution during image segmentation with iterated condition mode（ICM）, Metropolis-simulated annealing（MSA） based simulated annealing （SA） process were used in the current investigation. Results The global optimum solution were realized, and the proposed method can distinguish vessel as small as three voxels. The TOF-MRA data of Nanfang Hospital Imaging Center was used, which were collected from a 1.5 T GE MRI scanner with spatial resolution of 0.43 mm × 0.43 mm × 0.50 mm, raw data pixel size of 512 × 512 × 128 and actual size of 0.80 mm× 0.80 mm× 1.20 mm and 256 ×256 ×64. Each set of clinical data was compared using SA, ICM, MSA algorithm segmentation, and the results showed finite differences. The 15 iteration time consumption were 1 029 seconds, 463 seconds and 560 seconds. Conclusion It is demonstrated that the segmentation results of three-dimension simulated data display smaller global error with the model. Meanwhile, the segmentation of real MRA data also demonstrates good effects in comparison with the maximum intensity projection.