Automatic Generation of a Dynamic Anatomical Model for Use in Subsequent Radiological Image Interpretation

Abstract

Cranial anatomy described in texts depicts a part of hierarchy whose divisions reflect both artificial, sometimes historical, and physical composition and positioning. We have constructed a model of normal anatomy using a frame based representation. Each frame corresponds to a single structure in the hierarchy (or tree) and gives its composition in terms of smaller structures. Differentiation of features by properties takes place at a high level through classification by tissue type. The four classes used are: solids, deformables, fluids, cavities. The base level structures are described in detail for each slice in which they occur, using relative position, shape and size. This has been done on a coarse scale over a restricted range of slices in the first instance. This is a 3-dimensional representation from which an expectation of a 2-dimensional slice, obtained from a tomographic technique, can be generated. The base level features which occur in the specified slice range are taken from the anatomical model and used to construct a dynamic current world view. This is an initial hypothesis which can be altered in light of new information. The relevant parts of the rest of the hierarchy are then constructed above these features, considerably simplifying the anatomical model. Modality dependent knowledge is applied to the current world and the tissue type instantiated to the grey level values normally expected from that modality. Other attributes are then derived to evaluate the distinguishability of each of the base level features from its neighbours or beyond. Those which cannot be distinguished are merged. There are two types of merging which occur. (a) If neighbouring features are of the same tissue type then they are merged — they are parts of the same feature. (b) Either the modality or the subsequent interpretation information may determine the minimum grey level contrast, below which neighbouring features can be merged. Contrast and quality of gradient have been successfully used with an interpretation system to recognise the lateral ventricles and caudate nucleus, distinguishing the latter from the internal capsule.