Characterizing Myocardial Ischemia and Reperfusion Patterns with Hierarchical Manifold Learning

Abstract

We aim at better understanding the mechanisms of ischemia and reperfusion, in the context of acute myocardial infarction. For this purpose, imaging and in particular magnetic resonance imaging are of great value in the clinic, but the richness of the images is currently under exploited. In this paper, we propose to characterize myocardial ischemia and reperfusion patterns across a population beyond the scalar measurements used in the clinic. Specifically, we adapted representation learning techniques to not only characterize the population distribution in terms of scar and microvascular obstruction patterns, but also regarding the appearance of late gadolinium images which reflects tissue heterogeneity. To do so, we implemented a hierarchical manifold learning approach where the embedding from a higher-level content (the images) is guided by one from a lower-level content (the infarct and microvascular obstruction segmentations). We demonstrate its relevance on 1711 late gadolinium enhancement slices from 123 patients with acute ST-elevation myocardial infarction. We designed ways to balance the contribution of each level in the hierarchy, and quantify its impact on the overall distribution and on sample neighborhoods. We notably observe that the obtained latent space is a balanced contribution between the two levels of the hierarchy, and is more robust to challenging images subjected to artifacts or specific lesion patterns.