Each frame represents an individual anatomical feature and consists of a series of slots, and each slot has an associated value. The slots and their associated values provide an object centred description of the structure in question. The primary slot for each frame "defines" the structure being described. This slot's value is the anatomical name of the structure. Of the remaining slots, the most important is the "shape" slot. Unlike previous approaches to high level symbolic models, the shape slot contains an explicit lattice description for the object described in the frame. The shape description used for each structure is based on the Delaunay triangulation and its dual the Voronoi diagram, and is described in more detail in Robinson et al. . Input to this representation technique is in the form of candidate edge points without known connectivity. The Delaunay triangulation of a set of points connects points to their nearest neighbours. Each triangle is such that its circumcircle contains no points in its interior. The Voronoi diagram can be formed by joining the circumcircle centres of adjacent triangles, such that each Delaunay triangle side has a corresponding Voronoi diagram side perpendicular to it. However, the resulting Voronoi diagram sides may be offset from, and not necessarily pass through, their corresponding Delaunay triangle side. The Voronoi diagram sides form cells around each point enclosing the portion of the plane closer to that point than to any other. In the shape description used the Delaunay triangle sides are assumed to form a superset of the boundaries of the perceived objects, and the Voronoi diagram sides a superset of the skeletons of the perceived objects. Using a simple measure based on proximity, the correct subsets are retained and a dual boundary-skeleton shape representation is computed. This representation is then extended to produce a lattice description of each object in terms of its sub-parts. This whole process constitutes the shape representation module of the system.
We obtained the model shape for each object from the Talairach atlas as this provides typical cross-section views of the features that needed to be encoded in the model, all of which are presented in a 3D coordinate system. Each transverse view of the Talairach atlas was digitized, and the individual slices were manually aligned using the grids present on the digitized images to provide registration points. For each structure in the model, a series of points outlining it was manually chosen. These points were then used as input to the shape description described by Robinson et al . The resulting shape descriptions were then stored for each object. Only two dimensional representations of the shape of anatomical objects were acquired. The value of the shape slot in each frame is a reference to the file containing the description of the structure in question. This allows access to the shape information for any structure for a number of purposes including symbolic matching and visual inspection.
A number of other slots are also present in each frame. The exact number and type of slots coded varies according to the individual structure and class of structure being described but several slots are generic. Each frame has a slot describing the position of the centre of mass of the object being described. These positions were calculated from the individual objects' outlines (obtained from their shape representations) computed from the Talairach atlas. Each frame also has a slot stating the tissue type of the object it describes. This information can be used in conjunction with any known imaging information, such as MRI or CT acquisition parameters, to predict an approximate grey-level or range of grey-levels that the object is likely to have in the input image. Where applicable, frames also have a "function" slot. This slot, although not directly related to recognition, is information that can be extracted from the model and may be useful for other applications such as teaching.
Relationships between objects are coded as links between individual frames. However, these links are just a special type of slot called a "relation-slot". The value of relation slots must be another frame. This allows a "pointer" or "link" between frames. The properties of individual links are described in section 3.3. The information in each frame is essentially a LISP data structure, with each slot-value pair being an element in the data structure.