Regression and rejection of automated left ventricle boundary delineation

A boundary regression method has been studied to improve the performance of automated boundary delineation (ABD) from left ventriculograms. The regression transforms the original ABD boundary to a place in the boundary shape space where the coordinates are more like an LV boundary. The regression is generalized by the number of coordinates transformed. A rejection classifier is then designed to judge the reliability of the regressed ABD boundary. It takes the parameters derived from the original ABD boundary and the regressed boundary as the feature vector. The judgement is based on the difference between the two boundaries. The experiment results which were cross validated over a database of 375 studies are presented at the end of this paper. It shows that both the regression and rejection are helpful to improve the ABD performance