EXPERIMENTAL PLAN

Principal Investigator/Program Director Williams,Robert W.

AIM 2: Affined mapping into the MBL standard coordinate system

The main objective of this aim is to establish homology at a resolution of 225 mm among the brains stored in the MBL so that scientists using the MBL can readily retrieve similar planes of view from multiple animals by specifying the desired plane on the atlas. As part of this aim, we will develop a new version of our alignment toolbox that incorporates 2.5D registration, and we will release a new version of NeuroTerrain that will support simultaneous viewing of multiple brains.


Figure 4. 2.5 distance-based alignment. Images of sections from a rat brain were aligned to the blockface data of the same animal. Shown in A is the stack of outer contour from the set of 100 sections. The contour from the set of contour is shown in B prior to alignment, in C after section-to-section alignment with moments and in D after 2.5DBA. Some distortion in shape along the axis perpendicular to the horizontal plane can be observed in panel C relative to A. 2.5DBA improves the alignment appreciably. Color codes for distance from observer.

2.5-dimensional alignment

An obstacle in registration of histological material to a 3D atlas is that the sectional material is an ordered but unaligned 2D data set. When every section has been cut and stained, correlational rigid-body alignment (Toga 1993) using the section silhouette can be employed to first reconstruct the volume. 3DBA (Kozinska et al., 1997) can then be used to register the brain to the most suitable atlas. When gaps exist in the data set, as is the case in the MBL data, reconstruction using correlation is not possible. We will instead use a 2.5D extension of 3DBA (2.5DBA).

The distance-based alignment algorithm is available as a MatLab application <coe.drexel.edu/ICVC/Align/align11.html>. The latest version, Align 1.2 (not yet released), includes an implementation of 2.5D. Our tests of this approach have been confined to intra-animal alignment and phantoms where it does perform well (Fig. 4). For use in the setting of interanimal registration, the algorithm will be modified. The initial step in the registration process is section-to-section moments-based alignment. This achieves a fair degree of local registration accuracy. Next is an iterative alternation between 3D-3D and 2D-2D alignment. The 3D-3D alignment positions the volume as a whole relative to the atlas while the 2D-2D seeks the best in-plane rotation and translation of a given section to match the homologous atlas plane. In the present implementation, the 2D-2D alignment is unconstrained, and substantial inter-section shear may be introduced. To limit these problems and maintain a smooth section-to-section transition, we will impose inter-section coupling to penalize excursion from the initial position. The actual alignment is estimated to take less than 1 hour/brain and will require only 510 minutes of human inspection of automatic segmentation (using isodata thresholding) of the outer section contour.

Align 1.1 is now available on multiple platforms (IRIX, SOLARIS, Windows and MacOS). If proven successful, our extension of 2.5DBA will be released with an online manual and tutorial as we have done for Align 1.2.

Next Topic

• Aim 3: Segmentation of the MBL and extraction of quantitative traits.