Spatiotemporal Registration of Longitudinal Brain MRI Acquisitions by Identifying and Matching Multiresolution Triplet Pairs

In time-series analysis of longitudinal brain Magnetic Resonance Imaging (MRI) acquisitions, the first step is registration for any spatial misalignment in the acquired MRI volumes. This study demonstrates a ro-bust spatiotemporal registration technique by identifying multi resolution geometric triplets in a 3D space in an Euclidean reference axis for two MRI volumes of a patient undergoing ayurvedic medical intervention after an ischemic stroke. The desired triangular markers are those points in the fatty tissues of Tl MRI that are less prone to natural deformation over a few months, unlike the soft tissues in the brain that naturally deform fast and are studied by medical experts using brain MR. Here, we have used 3D spatial-frequency features of a Dual-Tree Complex Wavelet Transform (DTCWT) of the two longitudinal Tl brain MRI volumes to identify the triplets required for registration. These multiscale volumetric registration triplets are invariant to minor shifts and rotations due to the omission of down-sampling at the first decomposition level and interleaving samples at subsequent levels. These corresponding triplets of two Tl MRI volumes are matched by minimizing their Euclidean distances, and the optimal rotation required for their registration is estimated using a voting algorithm. The voted multiscale triplets first complete registration to rotations and then ensure registration to shifts.