Structural Informatics Group (SIG) logo
Home | Projects | Demos | Downloads | Publications | Local Info | About Us | New site

Artificial Intelligence

 

An active area of research in artificial intelligence is knowledge representation, on the hypothesis that intelligent systems act intelligently, in part because they have knowledge of the world. We are attempting to test that hypothesis in the domain of anatomy by developing a symbolic knowledge base in anatomy. This work is primarily being done by Cornelius Rosse and Onard Mejino, with help on the software side from Kraig Eno, Jeff Prothero and Jim Brinkley.

The work is driven by the need to develop an intelligent online anatomy information system in the Digital Anatomist Project. It is also an important component of the UW Human Brain Project as we develop a symbolic model of neuroanatomy that complements our spatial model.

Terminology

Our work in Symbolic Knowledge representation began with the development of a standard terminology for anatomical objects. We started with Neuronames, a neuroanatomical terminology developed by Doug Bowden and Richard Martin at the UW Primate Center. Neuronames is now part of the UMLS.

More recently, under a contract with NLM by Cornelius Rosse, we have developed a comprehensive set of terminology for thoracic anatomy that provides a unique name, plus associated synonyms, for all structures visible to 1 mm resolution, for a total of 14,916 terms associated with 8,763 concepts. (This number keeps changing as we add to the knowledge base). The terminology will be part of the 1998 release of the UMLS.

Semantic Network and class definitions

These terms have also been arranged in a semantic network (a very simple but useful knowledge representation scheme), in which the semantic classes are children of the UMLS semantic network, primarily the UMLS node, "Body Part, Organ, or Organ Component". Current links in the semantic network are of type "isa", "part of", "branch of" and "tributary of".

We have given definitions to the high level classes in the isa hierarchy.

The organization and motivational principles for this classification, plus an expanded set of definitions, are described in a JAMIA article [Rosse 1997].

Implementation

The semantic network and associated text definitions comprise the current implementation of our symbolic knowledge base. The terms and links are implemented as two tables in a relational database (Sybase and, on an experimental basis, msql). Authoring and end user clients access these tables by means of a standard relational database server, which can be either accessed directly, or indirectly via the knowledge server program, written in our version of Lisp called SLisp. The knowledge server provides a higher level query language to the clients, at the cost of reduced efficiency.

Further work

Now that the basic class structures have been defined, we are extending the terminology to include the abdomen, as an extension of Cornelius Rosse's contract with NLM. We are looking at more sophisticated knowledge represention schemes, and better knowledge acquisition tools. We are starting to extend the representation to include qualitative spatial relationships such as "anterior to". These extensions are being developed through collaboration with Linda Shapiro, Ira Kalet, and Pam Neal (a EE PhD student).

We are starting to test the representation by thinking of queries of the type expected by the anatomy information system, and asking whether the representation will be adequate to handle them. These queries will also help us develop inference methods for reasoning with the knowledge base, for interfacing with the spatial models, and for performing intelligent queries on the database.