Training sample reduction through model feature selection in anatomical model development

Model-based reasoning techniques complement traditional image processing and pattern recognition to compensate for the ambiguities inherent in medical imagery data. In model- based reasoning, higher-order image features, such as inferred surface boundaries, are matched against parametric models of anatomy. The instantiated model can be used to predict approximate locations of other anatomical image features for further segmentation processing. The inferential power of this approach is limited by the accuracy with which anatomical models capture population statistics. One possible problem with a model-based approach is that it requires a large training samples to develop models. We show that selection of quasi- invariant features greatly reduces the population sample required to accurately model population distribution of features. Quasi-invariant features include certain ratios, angles and other functions of directly measurable image features that constrain each others values. When one observable is measured in the image, the likely values and locations of others are relatively constrained. For example, the ratios of lengths of phalanges has small variance across the population of all humans compared to the variance of each bone length individually. An experiment is presented on a population of 90 radiographs that supports this approach to segmentation for hand anatomy measures.© (1992) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.