Detecting myocardial infarction using medial surfaces: LV statistical modelling challenge: Myocardial infarction

Abstract

The shape of the heart is known to undergo significant alteration as a result of myocardial infarction. We hypothesize that the thickness of the heart wall is an important variable in discriminating normal hearts from those with such defects. In the context of the present statistical modeling challenge, with meshes provided to describe the epicardium and endocardium at end diastole (ED) and end systole (ES), we model local heart wall thickness using a medial surface representation of fixed single sheet topology. Such a surface lies between the heart walls and the radius of the maximal inscribed disk at each point on it reveals heart wall thickness. We align the ED and ES medial surfaces to one another using the coherent point drift algorithm, and then align each registered pair to that of a reference heart, so that locations within each medial surface are in spatial correspondence with one another. We then treat the radius values at these corresponding medial surface point locations as inputs to a support vector machine. Our experiments yield a 96% correct detection rate on the 200 cases of labeled test data, demonstrating the promise of this approach.