one way at representing knowledge is as "production rules" which have five major properties [2] : a) the incorporation of human skills in conditional "if..then.." rules; b) an increase in skill proportional to the enlargement of the knowledge base; c) the ability to solve complex problems by selecting rules and combining the result; d) adaptive determination of the best sequence of rules to execute; e) explanation of their result by reversing the line of reasoning. Rules are good for representing knowledge which occurs as a large number of discrete facts.

"Structured objects" is the name given to any method of representation whose major format for holding information can be directly compared to the nodes and links of a graph. Structured objects hold information about objects and the relationships between objects. "Semantic networks" were the first such system to be developed [3]. In the semantic network model, objects, concepts, situations and actions are represented by nodes. The relationships between these nodes are represented by labelled arcs. The labels in the nodes gain their meaning from their connections [4]. Although there are no constraints in the naming of nodes and arcs, certain classes of links have developed as standards. For example the "IS-A" link is often used to convey the meaning of type. Descriptions of complex relationships between objects can be built up using this type of semantic network. Semantic nets can also be used to link attributes to objects. By combining the lattice structure that comes from the IS-A links with the attribute links, a network can be developed that will support the facility for property inheritance. That is, objects at the lower levels in the structure inherit properties explicitly associated with objects at the higher level. The major problem of semantic networks is the lack of theory of semantics. No fundamental distinction is made between different types of links, for example those that describe properties of objects, and those that make assertions, nor between classes of objects and individuals [5]. Niemann et al. [6] describe the development of system for the representation of knowledge based on the semantic network model. They show how semantic networks can be used for knowledge based image and speech understanding.

"Frames" [7] are an approach to structured object representation which allows more systematic grouping of information. They are intended to store stereotypical representations of different situations, objects and events. Individual frames correspond to the nodes in the network, and the relationships between the frames are the arcs in the network. Each individual frame stores information about a particular object or class of objects. It can be thought of as a record data structure consisting of a number of "slots", and associated with each slot there is a "value". The slots form a description of the object, with each slot-value pair corresponding to a common attribute of the object and its value or allowable range of values. Frames are arranged into hierarchies of classes via defined relations. Type hierarchies (constructed using "IS-A" or "A-KIND-OF" links) provide mechanisms for inheritance. One of the characteristics of frame systems is that information at the top of a class hierarchy is fixed, and as such can provide expectations about specific occurrences of values at individual frames. No slot values are ever left empty within a frame; rather they are filled with "default" values which are inherited from their ancestors. These default values form the stereotypical object, and are overwritten by values that better fit the more specific case. There are four methods for extracting information from a slot in a frame. Firstly if there is a specific value in a slot for a frame then this can be returned as it is assumed to be the most reliable and up-to-date value for the slot. Failing this, the system can either look at successive ancestors of the frame in question and attempt to inherit values which have been asserted into one of these frames (which is known as "value inheritance") or the system can look at the frame itself and its ancestors for procedures (known as "daemons") which specify how to calculate the required value (which is known as "if-needed inheritance"). If none of the above succeed in obtaining a value, then the system resorts to extracting the default values stored for the prototype object in question.

As a method for storing a high-level model of anatomical features for use in a computer vision system, frames are very attractive. They allow the generic features to be represented, the relationships between these features, and they also allow for specific instances of features (detected from input images) to be stored within the same framework. Using the inheritance features of frames it is then possible to guide processing by looking for specific arrangements of features within the input images.

A number of authors have used high-level models to aid in the analysis of medical images. Examples include the following: the use of high-level models to interpret CT scans [9][8]; The use of high-level models to interpret MR images [13][12][11][10]; and the use of voxel models to help interpret PET, SPECT and MR images respectively [16][15][14].

Hohne et al [17] have developed a model of brain anatomy for the teaching of medical students. In their model high level knowledge is linked to a voxel model of a particular segmented brain. Their approach to high level knowledge representation is similar to our own model which has been developed as an aid for brain image interpretation.


Fri Jul 15 14:54:31 EDT 1994