Brain Segmentation using ATP (Automatic Twice PAM) in Multi Diffusion Indices
Y-P . Chao, K-H . Cho, C-H . Yeh, S-P . Tsao, Chen Dy, Chen Jh · 2006
. All the procedures of animal experiment adhered to the guidelines for care and use of experimental animals of the lab animal center in National Yang-Ming University. Human brain images were acquired in a GE Healthcare Signa 1.5T Excite scanner by using spin echo EPI sequence with 252 diffusion-encoding directions, matrix size=128×128, slice number=4, voxel size=2.5×2.5×2.5 mm 3 , TR/TE = 2000/91.2 ms and bmax =3000 s mm -2 . Diffusion indices, FA, RA, RD, and Trace{D}, were calculated according to the standard formula [7, 8]. PAM was selected to replace conventional k-means algorithm for the medoid is less influenced by noise or other extreme values. Twice cluster processes, with better processing speed, were implemented to segment the DTI data in this study. Regions of interesting were divided into several 16×16 square regions to proceed first clustering. The four features, FA, RA, Trace, and RD, were normalized to 0-1 for calculating the related distance between pixels. 12 cluster medoids in each divided region generated from the first clustering were grouped up to decide the final cluster medoids, from which all pixels would be re-assigned. The number of clusters in the second clustering step depended on that of structures in the ROI and the location of slices. No initial information or input prior to processing step is needed in our method. All programs were developed on our own by using Borland C++ Builder 6 and OpenGL API.