Long term goal: Automatic screening of computerized tomography (CT) images of the head by i) the creation of a 3D coordinate System from the image file header, ii) use of low level image processing techniques to detect potential anatomical structures and iii) use a knowledge base system (KBS) to classify the structures from a 3D model.
Specific aims from this pilot study: i) Evaluate available KBSs and ii) investigate how variation in KBS construction affects recognition accuracy. Health relatedness: Automatic image recognition (MR) of CT images may improve diagnostic accuracy without significant cost increase. It would need a system which could classify patients as normal or abnormal which is the first major step in the diagnostic process. Abnormal images would be flagged prior to a radiologist's viewing. The radiologist not the computer would make the diagnosis. In essence the computer would be contributing to a double viewing method which has been shown with radiologists to improve diagnostic accuracy. This should decrease health costs and morbidity. Only 5 AIR systems of the head have been reported and four of these explicitly incorporeity knowledge in some form. However, the influence of varying the KBS components on recognition success has not been studied. Research design and methods: i) Potential KBSs will be evaluated for ability to a) interface to image data, b) represent data iconically, c) integration of semantic, spatial and grey level information, d) user interface and e) speed of operation. ii) 30 sets of adult axial CT images will be subjectively examined to identify image features to describe the maxillofacial region of the head. These image features will be detected by writing appropriate 'C' software. Their spatial relationships and distinguishing grey level properties will be stored in the KBS. Only bone not soft tissue will be studied to simplify the pilot. General anatomical spatial relationships will be manually entered into the KBS model. Testing different KBS structures: lO patient data sets (of up to 20 images per patient) will be repeatedly examined with variations in the KBS data and rules. The images will automatically be given english labels identifying anatomical points which are either true or false. Outcome: The alteration of KBS structure will be compared to rates of successful anatomical feature recognition.
