A knowledge-based system paradigm for automatic interpretation of CT scans

The interpretation of X-ray CT scans is a task which relies on specialized medical expertise, comprising anatomical, modality-dependent, non-visual and radiological knowledge. Most medical imaging techniques generate a single scan or sequence of two-dimensional scans. The radiologist's experience is gained by interpreting two-dimensional scans. The more complex three-dimensional anatomical knowledge becomes significant only when non-standard slice orientations are used. Hence, implicit in the radiologist's knowledge is the appearance of anatomical structures in standard two-dimensional planes, transverse, sagittal and coronal. That is, position with respect to both a coordinate reference system and other structures; intensity ranges for tissue types; contrast between structures; and size within the slices. Further to this, neurological landmarking is used to establish points of reference, i.e. more easily identifiable structures are first found and subsequent hypotheses are formed. With this in mind we have developed a knowledge-based system paradigm that partitions an image by applying the domain-dependent knowledge necessary (1) to set constraints on region-based segmentation and (2) to make explicit the expectation of the appearance of the anatomy under the imaging modality for use in the region grouping phase. This paradigm affords both expectation- and event-driven segmentation by representing grouping knowledge as production rules.