Data Driven Feature Learning for Representation of Myocardial BOLD MR Images

Cardiac phase-dependent variations of myocardial signal intensities in Cardiac Phase-resolved Blood-Oxygen-Level-Dependent (CP-BOLD) MRI can be exploited for the identification of ischemic territories. This technique requires segmentation to isolate the myocardium. However, spatio-temporal variations of BOLD contrast, prove challenging for existing automated myocardial segmentation techniques, because they were developed for acquisitions where contrast variations in the myocardium are minimal. Appropriate feature learning mechanisms are necessary to best represent appearance and texture in CP-BOLD data. Here we propose and validate a feature learning technique based on multiscale dictionary model that learns to sparsely represent effective patterns under healthy and ischemic conditions