Home Computer Science Computational Diffusion MRI: MICCAI Workshop, Athens, Greece, October 2016
Brain atlases [1, 2] capture the common features of image populations and play crucial roles in the processing and analysis of brain images. They are widely used for guiding brain tissue segmentation, normalization of images to a common space, and brain labeling with regions of interest. Unlike atlases of T-weighted images, diffusion MRI atlases afford additional white matter microstructural information that can be harnessed for tissue characterization and axonal tracing. To ensure that the microstructural information captured at each voxel location is properly encoded in the atlas, dedicated techniques are needed.
Atlas construction generally involves fusing a population of images that are registered to a common space. However, in practice, perfect registration is difficult, if not impossible. Averaging misaligned images to construct an atlas blurs structures and introduces artifacts. In diffusion MRI , the problem is even more challenging, since the alignment of gross anatomical structures does not necessarily guarantee the alignment of the microstructural information captured in each voxel. In this situation, it is unclear for example how signals characterizing fiber bundles of varying orientations, which can occur naturally across subjects, should be fused to form the atlas. Moreover, the commonly used simple averaging method is sensitive to outliers. For instance, if the distribution of signal profiles of singledirectional fiber bundles is contaminated with a small number of signal profiles of crossing fibers, simple averaging will result in a crossing profile, albeit with a small secondary peak. This outcome apparently is not representative of the majority.
In this paper, we propose a novel g-space patch-matching mechanism that is incorporated in a mean shift algorithm to seek the most probable signal at each point in g-space. Mean shift is a versatile non-parametric iterative algorithm that can be used for mode seeking . Instead of the mean, our method employs the mean shift algorithm to determine the mode of a distribution of signal profiles. The mean shift algorithm uses a kernel to measure the distance between signals. To increase robustness to noise, we measure the distance between signals using patches defined in the g-space. Patch matching is key to the success of many state-of-the-art denoising algorithms, such as non-local means . We perform patch matching in g-space with the help of azimuthal equidistant projection  and rotation invariant features . Experimental results confirm that our method yields diffusion atlases with cleaner fiber orientation distribution functions and less artifacts caused by intersubject fiber dispersion.
|< Prev||CONTENTS||Next >|