PCA regression for continuous estimation of head pose in PET/MR

With the availability of improved hardware and local point-spread function modelling, the presence of patient motion has become a major barrier to further improvements in the quality of PET images and their clinical efficacy. Although numerous approaches to compensate for patient motion have been proposed and are even commercially available, the additional hardware and extended setup time can preclude their routine clinical use. The MR modality on combined PET and MR (PET/MR) scanners can be used to correct motion with almost no additional setup time but currently must replace other MR acquisitions that may be required for clinic use. To overcome these problems, principal component analysis (PCA) and other data-driven techniques have been demonstrated to be able to reliably provide a signal related to patient motion based on raw PET data. Typically, these signals are used to split the PET acquisition into a discrete set of approximately motion-free time segments. This work introduces an approach where the PCA-signals are used as direct surrogates for the motion and regressed against rigid head motion parameters, enabling continuous pose estimation. A proof-of-concept is presented in which the approach is applied to upsample a low temporal resolution MR motion estimate. This proof-of-concept uses rapid echo planar imaging (EPI) data together with PET-derived motion signals. In a comparison of four techniques, nearest neighbour (NN) and linear temporal interpolation and linear and radial basis function (RBF) regression of pose against the PCA surrogate, we demonstrate that the model can be used to accurately interpolate pose continuously throughout the scan.