Development of a Model for use in Medical Image Interpretation

We present here progress on the Alvey MMI/134 project "Model based processing of radiological images". The radiological images we are dealing with are X-ray CT and NMR images of the head. Radiological interpretation of medical images obtained from any imaging modality, for example X-ray CT, relies on the fact that normal anatomy is predictable with respect to certain landmarks. The radiologist can then take into account variation between normal individuals and the effect of the imaging modality to create a flexible framework with fixed reference points to work from. We describe here a symbolic frame-based method of modelling 3D anatomy which allows 2D representations to be derived. This slice-wise representation is compatible with both the radiologist's view during interpretation and the images generated by the various imaging modalities. The types of radiological images^" produced when scanning the head are static 1 discrete 3-d volume data sets. This 3-d data set is composed of spatially contiguous and aligned 2-d discrete images. Each 2-d image (slice) from the sequence is completely defined by the slice projection and angle, an x,y co-ordinate system with respect to an anatomical co-ordinate reference system, a slice thickness (typically 5mm for X-ray CT) and grey-scale values. The appearance of anatomical tissue is dependent on the imaging modality used and on the situation the tissue is in. There is no real scope for changing the contrast in X-ray CT other than by use of contrast agents. Thus CT images from a particular scanner are fairly consistent but images may vary considerably between scanners. The problem of interpreting the appearance of tissue is exacerbated in NMR imaging due to the number of parameters associated with each tissue. The appearance of blood, for instance, can change dramatically when flow is present. When constructing a model!']'!"] for medical image interpretation assumptions can be made regarding the world being modelled. Firstly, the domain under consideration is well structured, with approximate prior constraints on location, shape of 3-d structures. Secondly, the domain can be explicitly described. Abnormality is described as being an absence, a deformation or a displacement of the normal. This will affect how the anatomy is perceived. Further to this an abnormal feature could be the presence of some additional