Automatic analysis of radiographic images: II. Software implementation and testing on bitewing radiographs

A prototype, completely automatic image-analysis system for dental bitewing radiographs was implemented in 'C' on a Unix workstation. A multiple layer hierarchical image model was created in software for each radiograph using production rule-based knowledge. The highest model layer formed an English-like description of anatomical structures including tooth type, interdental alveolar bone crest margins and spaces. A representation of spatial relationships between anatomical structures was formed, providing a method for comparing similar regions of a patient in serial, but not absolutely standardized, images. In 31 clinical radiographs, the system automatically indentified 93% (193) of all available interdental spaces which are the key image regions for model creation. Each image required 2 minutes of processing time. The reproducibility of finding 34 interdental spaces in five films revealed one error. Further work is required to improve the recognition rate of the interdental regions towards 100% and to develop automatic comparisons between areas of change (disease?) in serial films.