This paper describes the design of a computer system to teach the skills of visual categorisation. The initial domain of investigation is cardiac radiology , though the principles are generalisable to other domains. The Radiology Tutor is designed to carry out three types of task: to allow the student to browse through a database of digitised images; to form a teaching strategy that will provide a sequence of exemplar images for teaching; and to tutor about each radiographic image, selected either by the student or by the system, offering a critique of the student's interpretation and indicating abnormal features. Since human to human tutorial dialogues are primarily conducted at the level of anatomical features and their relationship to pathologies, the current system represents knowledge at this level. The system is not itself capable of interpreting the images; instead information about image regions and anatomical features is stored in frames associated with each image, with each pathology, and with the student's reported understanding of the image. Teaching about an image is scheduled by an agenda that provides a series of tutoring goals for the condition/action rules that control the tutorial interaction. A characterisation of images as points in a multi-dimensional feature space, pathologies as regions enclosing all the exemplar image points, and the student model as expanding regions enclosing the exemplars shown so far, provides a unified method of knowledge representation for the system.
This paper is not available online