Mammographic Knowledge Representation in Description Logic

Abstract

We present an advanced approach to representing knowledge about breast radiographs or mammograms which has advantages in terms of both usability and software engineering. The approach uses ontologies to create not merely a class hierarchy for a vocabulary but a full formal representation and, further, takes advantage of reasoning with description logic to provide application behaviour. The ontologies support a disjoint representation of graphical features and their interpretation in terms of medical findings. This separation of image features and medical findings allows the representation of different conceptual interpretations of the same graphical object, allowing different opinions of radiologists to be used in reasoning, which makes the approach useful for describing images to be used in computer-based learning and other applications. Three applications are discussed in detail: assessment of overlap in annotations, a conceptual consistency check in radiology training, and modelling temporal changes in parenchymal patterns. Reasoner usage, software testing, and implementation in Java are presented. The results show that, despite performance problems using the current implementations of reasoners, the description logic approach can be useful in practical applications.