Spatial Normalization

To make meaningful comparisons between images from different brains, extrinsic differences (position and orienta tion) must be removed and intrinsic differences (size and shape) minimized. A transformation process called spatial normalization (SN) is used to account for these differences by matching a set of brain features derived from a standard brain. Brain position, orientation, and size provide the minimal set of global spatial features for spatial normalization in three dimensions. When only global spatial features are used for the spatial transformation, the process is called global spatial normalization. Global SN is usually done using a nine-parameter affine transform with three parameters each for rotation, translation, and scaling [32,35,55,57]. One additional global spatial feature, skew, has been used for global SN with minimal improvement [43]. The global features of the brain carefully documented in the 1988 Talairach Atlas [32,49] are ideal for global SN, and this brain is used by the majority of brain mapping centers as the standard for spatial normalization. A brain image that conforms to the global spatial features of this standard brain is said to be Talairach spatially normalized and registered in Talairach space. Talairach space carries with it a Cartesian reference frame, and all measurements ( positions, distances, sizes, angles, and shapes) are made in this space. The most common usage is to report locations in the brain with x-y-z Talairach coordinates.

When greater detail is needed for feature matching, additional features and alternate transformation methods must be

Copyright © 2000 by Academic Press.

All rights of reproduction in any form reserved. 555

employed. The stereotactic transform of Talairach [48,49] is a classic example of a low degree-of-freedom regional method. It uses a piecewise linear fit to transform source images to match the brain in 12 regions of the Talairach atlas brain. The 12 regions are defined exteriorly by the 6 plane surfaces of the brain's bounding box and interiorly by 4 planes: coronal through the anterior commissure (AC), coronal through the posterior commissure (PC), and sagittal and axial planes intersecting the AC-PC line (Fig. 1). In this scheme there are four regions anterior to the AC, four between the AC and PC, and four posterior to the PC, divided equally between left and right hemispheres. Following reorientation of the source brain to match the atlas brain, corresponding internal and external planes of a source brain are identified. Internal planes of the source are transformed to match paired planes in the atlas, and proportional scaling is applied between parallel planes to stretch the brain to conform within each of the 12 regions [46]. An advantage of this method is that x-scaling can be different on the left and right, z-scaling can be different above and below the AC-PC line, and y-scaling can be different in the front, middle, and posterior regions of the brain. Talairach regional SN provides some additional feature matching capabilities when compared to Talairach global SN methods [57]. However, because of the large number of landmarks required,

Talairach regional SN cannot be directly applied in low-resolution images (PET and SPECT) where these anatomical features are not visible.

More powerful transform methods are necessary to extend feature matching to the level of detail in high-resolution MR brain images [5,6,11,26,31,34,57]. Such methods often employ large deformation fields to perform spatial normalization. These high degree-of-freedom methods cannot directly match the 1988 Talairach atlas since their feature-matching algorithms require identical MR image characteristics for both source and target brains. The target brain image is globally spatial normalized to overcome this limitation. The modified target image is in Talairach space, and locations within brain images regionally transformed using the modified target also reference the Talairach space. The significance of global SN of the target brain needs to be emphasized, since it is the global component of regional SN that provides the frame of reference, i.e., the transform to Talairach space. Although regional SN components can provide superior regional feature matching, they should not alter this reference frame. This scheme provides Talairach space compatibility for spatial normalization methods ranging from nine-parameter global methods to high degree-of-freedom deformation-field methods.

FIGURE 1 Twelve regions of the 1988 Talairach atlas. The outer boundaries define the bounding box for the brain. Talairach coordinates are given for the AC and PC and bounding box coordinates at the intersection of the x, y, and z axes. The orientation of the brain within this coordinate system is illustrated in Figs 3 and 5. See also Plate 77.

FIGURE 1 Twelve regions of the 1988 Talairach atlas. The outer boundaries define the bounding box for the brain. Talairach coordinates are given for the AC and PC and bounding box coordinates at the intersection of the x, y, and z axes. The orientation of the brain within this coordinate system is illustrated in Figs 3 and 5. See also Plate 77.

Sourcc

Spat ml Normalization Algorithm

1, Sel Transformed at Source.

2, Bxiruci features (Fs).

3, Compare features (AF),

4, Transform source to match features,

5, Repeat 2-4 until done.

Reverse Testicular Atrophy

FIGURE 2 The schematic plan for spatial normalization of brain images.

FIGURE 2 The schematic plan for spatial normalization of brain images.

0 0

Post a comment