Data-driven feature learning for myocardial registration and segmentation

This chapter discusses novel techniques for the tasks of segmentation and registration separately and jointly. In particular, the feature learning is tested on cardiac phase-resolved blood oxygen-level-dependent (CP-BOLD) MR images, which is a new contrast agent- and stress-free imaging technique for the assessment of myocardial ischemia at rest. CP-BOLD MRI introduces varying contrast in medical image analysis applications. Therefore, establishing voxel to voxel correspondences throughout the cardiac sequence, an inevitable component of statistical analysis of these images remains challenging. Furthermore, medical background and specific segmentation difficulties associated to these images are present. Alongside with the inconsistency in myocardial intensity patterns, the changes in myocardial shape due to the heart's motion lead to low registration performance for state-of-the-art methods.

The problem of low accuracy can be explained by the lack of distinguishable features in CP-BOLD and inappropriate metric definitions in current intensity-based registration and segmentation frameworks. In this chapter, sparse representations, which are defined by a discriminative dictionary learning approach, are used to improve myocardial segmentation and registration. Initially appearance information is combined with Gabor and HOG features in a dictionary learning framework to sparsely represent features in a low-dimensional space. For registering, the cardiac sequence a new similarity metric is proposed utilizing the sparse representations. Also, an alternating optimization scheme for dictionary learning-based feature representations is proposed using the sparse coefficients and dictionary residuals. The superior performance of the dictionary-based descriptors are showcased with several experimental results.