Thesis Abstract: Dr Darryl N. Davis
Wolfson Image Analysis Unit, Faculty of Medicine, Victoria University of Manchester, 1991

KNOWLEDGE-BASED SYSTEMS FOR MEDICAL IMAGE INTERPRETATION
 

The aesthetics of cranial and facial form have, throughout the ages, been based on a number of criteria. In modern orthodontic practice great reliance is placed on systematic and objective methods of characterising craniofacial forms, using measurements based on both hard and soft tissue landmarks. Lateral skull X-ray images are routinely used in cephalometric analysis to provide quantitative measurements useful to clinical orthodontists. This thesis describes research work aimed at automating cephalometric analysis.

Initial investigations into the processing and segmentation of digital cephalometric images established a background for the project. A simple image interpretation system was devised as means of determining the problems associated with the automation of the cephalometric analysis. These initial investigations, and a study of the relevant literature, suggested that a flexible and adaptive vision system was required to successfully interpret digitised lateral skull radiographs, or indeed any image from a difficult (biomedical) domain. It is argued that a model and knowledge-based methodology offered the best approach in satisfying these requirements. A rule-based segmentation system, making use of an image appearance model, was developed to extract image features from grey-level images.

Complex image features and cephalometric landmarks are constructed from these segmented component features. A predictive model, defining picture structure, was developed, allowing location hypotheses to be made for image features. The underlying structure of the location model provided the basis for a geometric constraint model of use in discriminating between image feature candidates. An automatic image interpretation system was developed to find, in a given order, a given set of image features and cephalometric landmarks. This system was found to lack the means to organise and successfully complete image interpretation tasks, but paved the way for a blackboard system for cephalometry. This system used a hierarchical approach to task organisation, with individual knowledge sources grouped according to function and the individual stages of the adopted image interpretation cycle.

        A detailed analysis of the results from applying the image segmentation and the two interpretation systems is described. The blackboard system demonstrated an identification rate comparable to the best previous solution. It was also capable of interpreting misaligned and biologically variant cephalometric images, which caused problems for earlier systems. The main body of the thesis concludes with a presentation of some ways of improving and extending the blackboard system.