A new shape-based framework for the left ventricle wall segmentation from cardiac first-pass perfusion mri

We propose a shape-based approach for the segmentation of the left ventricle (LV) wall on cardiac first-pass magnetic resonance imaging (FP-MRI) using level sets. To reduce the variabilities of the LV wall in FP-MRI, it is first imperative to co-align the time series images to account for the global and local motions of the heart. Therefore, we developed a two-step registration methodology that includes an affine-based registration followed by a local B-splines based alignment to maximize a similarity function that accounts for the first- and second-order normalized mutual information (NMI). Additionally, myocardial signal intensity varies with the agent transit, which makes it difficult to control the level set evolution using image intensities alone. Thus, we constrained the level set evolution using three features: a weighted probabilistic shape prior, the first-order pixel-wise image intensities, and a second-order Markov-Gibbs random field (MGRF) spatial interaction model. We tested our approach on 24 data sets in 8 infarction patients using the Dice similarity coefficient (DSC), comparing our approach to other shape-based segmentation approaches. We also tested the performance of our segmentation approach using the receiver operating characteristics (ROC). Our approach achieved a mean DSC value of 0.910±0.037 compared to other shape-based methods that achieved 0.862±0.045 and 0.844±0.047. Finally, the ROC analysis for our segmentation method showed the best performance, with area under the ROC curve of 0.92, while that for intensity showed the worst performance, with area under the ROC curve of 0.69.