This paper describes the design and computer implementation of a high level symbolic model of the human brain. Unlike most other models of the brain it contains a part hierarchy and descriptions of the spatial adjacency relationships between anatomical structures. Individual structures are represented by frames. These frames contain object centred descriptions of the structures including a shape representation. The knowledge that forms the basis of the model was extracted primarily from the Talairach neuroanatomical atlas with additional information from clinical neurologists and text books.
The model can be considered as comprising three co-existing graphs of relationships between anatomical structures. These are a hierarchy of parts and sub-parts of structures, a graph of spatial adjacency relationships, and an inheritance network for structures within the model and objects extracted from the input images.
We have also presented a mechanism for keeping track of potential multiple matches of objects extracted from the input images to anatomical structures held within the model itself. This mechanism is very similar to the "assumption-based truth maintenance system" (ATMS) approaches advocated by a number of different authors in other domains. The combination of the generated partial solutions with the structured and easily modifiable nature of the shape representation means that it is potentially possible to modify the partial solutions following further processing. However, if a large number of potential image objects are passed to the model environment for matching then there may well be a combinatorial problem with the merging/poisoning of the generated viewpoints. This type of high-level symbolic model has great potential not only in automated recognition of objects from medical images, but also as a teaching aid. The model of Hohne and co-workers [17] has been specifically designed for this purpose. Our model does not have the user interface to allow it to be used in this way, although the type of information stored means that it could be possible to develop it.
The model needs to be extended before it can be used efficiently and with success. A voxel representation of each object should be attached (as a "slot") to each node, providing a volume description of individual structures. This would have the effect of combining both the "voxel model" approaches proposed by some authors with the primarily symbolic approach that we propose.
A more systematic approach to the coding of spatial relationships also needs to be accomplished. There is a great deal of ambiguity and difficulty in computing the actual spatial relationships of complex shapes in both two and three dimensions, and there are linguistic problems in what we actually mean by left-of etc. Internally consistent definitions are paramount. These difficulties need to be tackled. Even so, the framework developed here provides the mechanisms for incorporating more complex and robust approaches.
Probably the biggest single improvement to the model would be its extension to three dimensional shape descriptions. This requires that a robust approach to the three dimensional extraction and representation of shape be developed. The use of the Delaunay-Voronoi dual could be extended to accommodate this [22]. It would also require that more knowledge is acquired about the three dimensional shape of the anatomical structures being represented, and also about the coding of spatial relationships between objects in three dimensions. Currently only information from a single plane of the Talairach atlas has been used to obtain the model information (the transverse plane) and information from the other planes and other sources will have to be incorporated. This would aid in the production of three dimensional representations of the anatomical structures, and consequently mean that a greater number of images could be analyzed using these methods.
The matching process described in the example given in this paper is only capable of providing a score for a match between two completely segmented objects. This is a significant limitation which means that a successful labelling is highly dependent on a good initial segmentation. A more complex matching algorithm based on the properties of the individual objects' shape representation is required which can allow partial matches between image objects and model objects. Such a matching process would necessitate a more sophisticated control process than the one currently implemented.