The general goals of the High Resolution Brain Atlas project are 1) to construct atlases of the brain of the mouse, an animal that is becoming increasingly important for neuroscience, particularly because of its position as the mammal of choice in genetic research, and 2) to use the brain of this small, relatively inexpensive and readily available animal as the test case on which to work out issues of capture and integration of multidisciplinary data, of 3D reconstruction from serial slices, and of data dissemination, in the expectation that the lessons learned will become applicable directly to the much larger normal and diseased human brain. The work accomplished and the anticipated future directions are enumerated below.

Creation of a 2D Digital Atlas of the Mouse Brain

Our first goal, now essentially achieved, has been to create a full 2D digital atlas of a two-month-old female C57BL/6J mouse brain at 10 micron voxel resolution. The atlas is based on a complete set of coronal sections stained alternately for cells and for myelin, some of which were used in the paper atlas of Sidman, Angevine and Taber Pierce (1971). The procedures for data capture, for adjustment of staining intensity and hue among the sections, for alignment of sections, and for nomenclature of gray and white matter structures (annotations) are described here. For ease of calling up images over the Internet at today's limited band width, the atlas has been divided arbitrarily into eleven broad regions each of 1000 microns length-The 1st through the olfactory bulb and the 11th through the caudal half of the cerebellum, each region composed of 25 Nissl-stained sections paired with 25 myelin-stained sections. A 12th region of 440 microns contains the caudalmost part of the brain, the low medulla, and is composed of the final 11 Nissl-stained sections paired with 11 myelin-stained sections. Also, the 2D atlas has a 13th region containing as representative pair of Nissl- and myelin-stained sections through each of the 33 spinal cord segments. For convenience of access, a pair of sections at 200 micron intervals through the entire brain is presented in the 2D atlas. In addition, composite images were constructed by mapping a cell (Nissl-stained) image onto the adjacent myelin image. An additional set of paired images is included with labels and leader lines added ("annotated"), representing the coronal sections at 200 micron intervals, plus additional sections that contain small structures which lie entirely between the 200 micron increments. Thus, every named structure can be visualized. The nomenclature matches that in the paper atlas referred to above, with some additional or alternative names from more recent publications. The user has the option of choosing English or Latin nomenclature.

All histological processing methods introduce distortions of brain tissue. For some purposes, such as theoretical modelling, absolute measurement of cell and tissue dimensions, strain comparisons and analysis of the actions of genes influencing strain differences, and comparative neuroanatomy, it is important to warp the histological images to some absolute "gold standard." The closest we come today to a "gold standard" is a set of magnetic resonance images (MRI) obtained over many hours at high magnetic strength, first on a fixed mouse head and then on the same brain, carefully dissected from the skull. Absolute dimensions are recorded, and all images are in alignment. Warping of our histological images to an MRI series from brains of mice of the same strain, age and sex are currently in progress. For other purposes, the unwarped images may prove to be of greater value, particularly for the overlaying of images obtained with other stains onto the "canonical" atlas's cell- and myelin-stained images. Among the examples that come to mind are Golgi and corresponding fluorescent images, immunohistochemical images, silver degeneration and other axon-tracing images, in situ hybridizations for gene expression, and pathological images. One could argue that a separate atlas should be prepared by every common histological preparative method (embedding in paraffin, glycol methacrylate, epoxy resin, or celloidin, cryostat or vibratome sections, etc.), but as a practical compromise, we propose to supplement the present 10 micron atlas of a celloidin-embedded brain with 3 micron and 1 micron atlases based on differently processed specimens, with and without image warping to the MRI standard (please see below).

Mouse Brain 3D Atlas and Volume Visualization

Another immediate goal is to create a 3D voxel atlas of the C57BL/6J mouse brain from the present 2D section images. The critical factors are alignment of section images and matching of stain intensities, because their precision controls the efficiency and accuracy of 3D segmentation. The 10 micron images have been aligned semiautomatically with newly written VoxelMath software, followed by small further manual adjustments using pixel-based visual landmarks. Full cell- and myelin-stained mouse brains have been constructed. Further refinement of the alignment of serial images is anticipated, and should precede final segmentation of gray and white matter structures. Differences in staining intensity from section to section may be hardly noticable when one studies sections in the standard manner, but the differences become very important during 3D reconstruction. Precise matching of stain intensities has proved to be unexpectedly difficult because, as we have documented, the light, medium, and heavily stained tissue components do not remap on the same linear scale. Not only do the positions of the stain intensity histogram curves differ from section to section, but even their shapes differ. Moderate success has been obtained by plotting the histogram of voxel intensities in a relatively large 3D sample of brainstem, and then mapping the entire series of section images to the shape and position of that sample's histogram. We are working with texture mapping algorithms to refine further the smoothing out of the staining differences, but even now we are able to automate the 3D segmentation of selected gray and white matter components, and expect to be adding the results to the atlas over the coming months. The segmented volumes will then be connected via an intuitively easy user interface to the informatics database of the High Resolution Brain Atlas.

Data Dissemination

Our aim is that the images be searchable through an informatics database by a query that addresses either the images themselves or the names on a hierarchical structure list. The database and the query range will be expanded gradually to link other data from neuroanatomy, chemistry, pharmacology, physiology, psychology, genetics and pathology. All images, including those in between the displayed images at 200 micron intervals, will be distributed via this website on demand by the user. We are setting up procedures and protocols to accomplish the delivery, and ask any user with an immediate need for some subset of the section images, or for atlas images resectioned electronically at an orientation matching the user's own specimen(s), to send mail to Richard L. Sidman.

From research of the past ~100 years, it is clear that maps are needed at different resolutions to provide different kinds of information. Many aspects of structure delineation in gray and white matter regions are best done at 10 micron voxel resolution, whereas cell classification and texture-based segmentation appear to be best accomplished at 3 microns, axon trajectories at 1 micron, and synaptic organization in, say, retinal plexiform layers, glomeruli in cerebellum and elsewhere, and in cortical sensory barrels and columns are resolvable only by electron microscopy at < 0.1 micron. The map we propose to develop at 1 micron resolution will meet the objective we and others have contemplated, of plotting for the first time the detailed trajectories of white matter tracts throughout the brainÑeven, for the larger caliber axons, down to the level of individual myelinated fibers. Such a refined anatomy in the mouse brain, especially cortico-cortical white matter anatomy, will likely prove to be extrapolatable to other mammals, and should delineate the fundamental white matter organization in even the human cerebrum. Such a body of information is essential for interpreting human data obtained by fMRI and related methods.

To plot such detailed anatomy manually would be an immense undertaking. We anticipate that automated and accurate segmentation will be feasible with thinner section images captured at high resolution. It will become realistic to trace trajectories of the larger myelinated axons for long distances through the brain, and to store data on how constant are the relationships among axons over long distances, especially in systems where we know from physiology that topographic representation is preserved across synapses. Cell size, shape and packing density, automatically recorded in the segmented 3D gray matter "clouds," should provide new useful and objective criteria for classification of structural units in the brain in supplementation of input/output relationships and chemical/genetic properties. The more complete the neuroanatomical delineation of features in the mouse brain, the better will neuroscientists be able to map new and old research data onto the atlas templates.