Project 1: The Mouse Brain Library Project 2: Internet Microscopy (iScope) Project 3: Neurocartographer and Segmentation of the MBL Project 4: The Neurogenetics Tool Box



























 

 

 

 

 

 

 

 

 

 

 

 

  

 

  
 

RESEARCH PLAN

  
 

Principal Investigator/Program Director Williams,Robert W.

 
 

Status of current research efforts

A prominent concept in neuroimaging is the use of anatomical templates positioned in a standard coordinate system. Experimental data are aligned into this coordinate system and consequently inherit labels for numerous neuroanatomical compartments. In addition to establishing homology at the level of anatomical structures, coordinates can be used directly to address equivalent points across brains.  A number of coordinate systems have been developed for human neuroimaging. The most commonly used is defined by the Talairach atlas, which was constructed from histological sections derived from postmortem material (Talairach and Tournoux 1988). Another system was defined by 3D MRI (Evans et al. 1994). In nonhuman research, the closest equivalent to a standard system consists of stereotaxic coordinates defined by their positional relation to bony landmarks (Slotnick and Brown 1980; Paxinos et al. 1985), but these are ill suited for registration of brains to an anatomical template because the relation between the brain and the cranium is lost during histological procedures. Furthermore, even if the relation could be recovered, the variability of the skull is often substantial and does not necessarily support accurate template placement. Brain atlases are more suitable for this purpose.

There are a number of 2D digital rodent brain atlases (Toga et al. 1989; Bloom et al. 1990; Nissanov and Bertrand 1998a). To truly model the brain and to accommodate arbitrary planes of sectioning, however, 3D atlases are needed. Such atlases have been constructed using MRI or blockface imaging for a number of species (Black et al. 1997; Cannestra et al. 1997; Ghosh et al. 1994; Toga et al. 1994, 1995), but these are not ideal in the setting of the MBL. As the underpinning for a coordinate system, atlases serve two purposes. First, their demarcation into anatomical regions defines a standard template to be used in parcellation of experimental material. Second, features on the atlas, either geometric or gray value, drive the registration process. MRI and blockface imaging fail to provide the resolution and the contrast needed. Computerized 3D atlases from stained sections, a more suitable basis for an MBL atlas, have been generated previously for the rat (Nissanov and Bertrand 1998b; Nissanov et al., submitted) and the mouse (Celio et al. 1998; Davidson et al. 1997; Reed et al. 1999). During the past few years, we have developed progressively more effective tactics for atlas construction. We have used these tactics to construct a 2D rat brain atlas (Nissanov and Bertrand 1998a), a 42-mm isotropic 3D rat brain atlas (Nissanov et al., submitted; Nissanov and Bertrand 1998b), and a 20-mm isotropic 3D atlas of C57BL/6J (Reed et al. 1999). In ongoing studies we are developing methods to construct atlases from celloidin-embedded tissue. Once constructed, our new atlases will guide registration of MBL data into the standard coordinate system. Two important considerations are the types of information used for this alignment and the classes of transformations used to register sections. Information for alignment can be either geometrical (model) or image intensitybased. In geometrical alignment, points, lines, surfaces, or geometrical invariants are abstracted, automatically or manually, from experimental images. These are matched to homologous geometric objects in the atlas coordinate system. In image intensitybased methods, a global disparity between the experimental and atlas image intensities is computed, and alignment is adjusted to minimize this disparity. Classes of transformation for alignment range from geometrical linear mappings (e.g., rigid-body or affine transformations) to complex nonlinear mappings

In intensity-based procedures, gray value disparity measures such as squared differences or correlation are useful in matching images with the same modality (Hibbard and Hawkins 1988; Collins et al. 1994; Thevenaz and Unser 1995). In the setting of histological material, a given stain constitutes a modality. More sophisticated and flexible disparity measures such as conditional variance (Woods et al. 1993) and mutual entropy (Wells et al. 1996; Thevenaz and Unser 1996; Maes et al. 1997) have been found effective for both intra- and intermodality registration. Many investigators are using the programs AIR (automatic image registration) from the University of California at Los Angeles  (Woods et al. 1998 a, b) and ANIMAL (automatic nonlinear registration) from the Montreal Neurological Institute (Collins and Evans 1996). AIR can perform both linear (rigid body, affine) and nonlinear (polynomial) alignment with a choice of three gray value comparison functions, allowing inter- and intramodality alignment. ANIMAL is designed for intramodality intersubject alignment; it incorporates a powerful nonlinear spline alignment model.

Geometric-based methods rely on prior image segmentation of fiducials. Once obtained, the fiducial registration points can be used to warp brains into a standard coordinate system. The class of transformations should be matched to the character of the matching problems; for example, rigid-body transformations are suitable, up to a point, in 3-D reconstruction (Hibbard et al. 1987) and multimodality images from the same subject (Pellizari et al. 1989), and the affine transformation is the classical method for intersubject alignment (Collins et al. 1994; Fox et al. 1994). A greater range of intrasubject variability can be accommodated with nonlinear transformations, which include polynomials (Friston et al. 1995; Woods et al. 1998 a, b), product splines (Collins et al. 1995), and transformations based on continuum mechanics (Bajcsy et al. 1983; Christensen et al. 1996; Kim et al. 1997; Gabrani and Tretiak 1999). 

For rodent brains, rigid-body alignment is surprisingly accurate even in the setting of interanimal alignment (Ozturk et al, submitted; see Appendix). We have employed 3D distance-based alignment to register brains using the outer surface as reference. Misregistration of the internal structures was found to be 193.5 mm (rms). Improvement will require nonlinear transformation guided by both the outer brain surface and internal fiducials. We propose to use spline interpolation for this purpose. Spline functions may be designed with the requisite degree of differentiability and are stable to local perturbation. The two general spline methods for multidimensional interpolation are (tensor) product splines (de Boor 1978) and radial splines (Duchon 1977; Meniguet 1979; Bookstein 1989; Gabrani and Tretiak 1999). Fiducial structures in brains occur at irregular locations, and radial spline interpolation is well adapted to such data. Product splines have certain computational advantages and have been used for brain registration by Collins et al. (1994). The key issue in constructing transformation functions is accuracy, which improves as more fiducial points are used and also depends on the smoothness of the function being interpolated. To reduce time and effort, it is desirable to use the fewest fiducial points possible. A commonsense resolution to this problem is to use more fiducial points in regions of high variability.

Although fiducial points are very easy to use, they are relatively hard to locate. Surfaces (or outlines in 2D images) are now commonly employed and have been the basis of many successful algorithms (Amit 1997; Borgefors 1988; Pellizari et al. 1989; Besl and McKay 1992; Jiang et al. 1992; Huang and Cohen 1996; Kozinska et al. 1997; Thompson et al. 1997; Cohen el al. 1998; Ibrahim and Cohen 1998; Gabrani and Tretiak 1999). The method we employ for computing transformations (described by Gabrani and Tretiak 1999) can use both types of data. Spline theory is based on defining an energy functional, typically the integral of the square of a derivative of the interpolating function. A spline interpolator is the function that minimizes the integral of the energy. For example, cubic splines minimize the integral of the square of the second derivative of the interpolation function, and Bookstein's (two-dimensional) thin-plate splines minimize the square of a Laplacian function. When we use surfaces for spline alignment, points from the surface of one object must be transformed to the surface of the second object. The location of the point on the second surface is not known but is found by minimizing the energy functional. The method is directly applicable to heterogeneous data consisting of fiducial surfaces, lines, and points.

No matter how the brains are aligned, the end result is immensely valuable. In addition to its use in segmentation, brain alignment has been employed to detect abnormalities (Thompson et al. 1997; Thompson and Toga 1997), for surgical planning (Levy et al. 1997; St.-Jean et al. 1998), and, most relevant to the present application, for database retrieval (Fox et al. 1994). For our purpose, alignment brings the brains into the coordinate system. Neuroscientists will then be able to access corresponding points across subjects by simply referring to a coordinate in the standardized system.

 

 
   
   
   
 

Next Topic