Ardekani B A, Braun M, Hutton B F, Kanno I, Iida H
Department of Applied Physics, University of Technology, Sydney, Broadway, Australia.
J Comput Assist Tomogr. 1995 Jul-Aug;19(4):615-23. doi: 10.1097/00004728-199507000-00022.
A fully automatic multimodality image registration algorithm is presented. The method is primarily designed for 3D registration of MR and PET images of the brain. However, it has also been successfully applied to CT-PET, MR-CT, and MR-SPECT registrations.
The head contour is detected on the MR image using a gradient threshold method. The head region in the MR image is then segmented into a set of connected components using the K-means clustering algorithm. When the two image sets are registered, the segmentation of the MR image indirectly generates a segmentation of the PET image. The best registration is taken to be the one that optimizes the segmentation induced on the PET image. In this article, the K-means minimum variance criterion is used as a cost function, and the optimization is performed using the method of coordinate descent.
The algorithm was tested on 80 H2 15O PET and MR image pairs from 10 subjects. Qualitatively correct results were obtained in all cases. With use of external markers visible in both image modalities, the average registration error was estimated to be < 3 mm.
The algorithm presented in this article requires no user interaction and can be applied to a wide range of registration problems. Quantitative and qualitative evaluations of the algorithm indicate a high degree of accuracy.
提出一种全自动多模态图像配准算法。该方法主要设计用于脑部磁共振成像（MR）和正电子发射断层扫描（PET）图像的三维配准。然而，它也已成功应用于CT-PET、MR-CT和MR-SPECT配准。
使用梯度阈值法在MR图像上检测头部轮廓。然后使用K均值聚类算法将MR图像中的头部区域分割成一组连通分量。当两个图像集配准时，MR图像的分割间接生成PET图像的分割。最佳配准被认为是优化PET图像上诱导分割的配准。在本文中，K均值最小方差准则用作代价函数，并使用坐标下降法进行优化。
该算法在来自10名受试者的80对H2 15O PET和MR图像上进行了测试。在所有情况下都获得了定性正确的结果。使用在两种图像模态中都可见的外部标记，估计平均配准误差小于3毫米。
本文提出的算法无需用户交互，可应用于广泛的配准问题。对该算法的定量和定性评估表明其具有高度准确性。
