Tracking Endocardial Boundary and Motion via Graph Cut Distribution Matching and Multiple Model Filtering

Tracking the left ventricular (LV) endocardial boundary and motion from cardiac magnetic resonance (MR) images is difficult because of low contrast and photometric similarities between the heart wall and papillary muscles within the LV cavity. This study investigates the problem via Graph Cut Distribution Matching (GCDM) and Interacting Multiple Model (IMM) smoothing. GCDM yields initial frame segmentations by keeping the same photometric/geometric distribution of the cavity over cardiac cycles, whereas IMM constrains the results with prior knowledge of temporal consistency. Incorporation of prior knowledge that characterizes the dynamic behavior of the LV enhances the accuracy of both motion estimation and segmentation. However, accurately characterizing the behavior using a single Markovian model is not sufficient due to substantial variability in heart motion. Moreover, dynamic behaviors of normal and abnormal hearts are very different. This study introduces multiple models, each corresponding to a different phase of the LV dynamics. The IMM, an effective estimation algorithm for Markovian switching systems, yields the state estimate of endocardial points as well as the model probability that indicates the most-likely model. The proposed method is evaluated quantitatively by comparison with independent manual segmentations over 2280 images acquired from 20 subjects, which demonstrated competitive results in comparisons with a recent method.