IMAGING METHODOLOGY - Notes Revised Motion Estimation Algorithm for PROPELLER MRI James G. Pipe, 1 * Wende N. Gibbs, 1 Zhiqiang Li, 1 John P. Karis, 1 Michael Schar, 2 and Nicholas R. Zwart 1 Purpose: To introduce a new algorithm for estimating data shifts (used for both rotation and translation estimates) for motion-corrected PROPELLER MRI. The method estimates shifts for all blades jointly, emphasizing blade-pair correlations that are both strong and more robust to noise. Theory and Methods: The heads of three volunteers were scanned using a PROPELLER acquisition while they exhibited various amounts of motion. All data were reconstructed twice, using motion estimates from the original and new algorithm. Two radiologists independently and blindly compared 216 image pairs from these scans, ranking the left image as sub- stantially better or worse than, slightly better or worse than, or equivalent to the right image. Results: In the aggregate of 432 scores, the new method was judged substantially better than the old method 11 times, and was never judged substantially worse. Conclusion: The new algorithm compared favorably with the old in its ability to estimate bulk motion in a limited study of volunteer motion. A larger study of patients is planned for future work. Magn Reson Med 72:430–437, 2014. V C 2013 Wiley Periodicals, Inc. Key words: MRI; motion; PROPELLER INTRODUCTION Due to the relatively long scan times in MRI, patient motion often results in image artifacts. In this context, one can approximate many structures (e.g., the brain) as rigid bodies, with three degrees of freedom for both rotation and translation applied to the entire volume in a time-varying manner. For two-dimensional (2D) rapid acquisition with refocused echoes (RARE, also called turbo spin echo [TSE] and fast spin echo [FSE]) imaging, the PROPELLER method (1) has been developed to estimate and remove the effects of in-plane rigid body motion, while mitigating the effects of through-plane motions with the use of data rejection. PROPELLER has been shown to be quite effective in mitigating patient motion (2,3), but it is not error-free. Inaccurate estimates of motion can lead to corruption of otherwise motion-free data sets, as well as result in non- optimal correction of motion-corrupted data sets. Subse- quent work (4,5) has analyzed the effect of various parameters and minor adjustments on the performance of original PROPELLER reconstruction algorithm. Based in part on the benefits of wider blades, Tamhane et al. (6) proposed undersampling blades in the phase encod- ing to make them wider, and combining this with itera- tive reconstruction. Nehrke et al (7) proposed using image-based registration based on Gauss-Newton optimi- zation for motion correction. Holmes et al (8) showed that, in the case where long echo trains are difficult to collect (e.g., T1-FLAIR PROPELLER), external calibration data can be useful to facilitate acceleration with parallel imaging to widen the blades. The original PROPELLER motion estimates are obtained separately for rotation and translation, but in both cases, the motion is framed as a simple shift estimate (in the angular direction for rotation after gridding k-space data to polar coordinates, or directly in x-y space for transla- tion). The algorithm successively maximizes data correla- tion between each blade and some reference blade to estimate the respective shifts. Two components of this shift estimation algorithm were identified as potential sources for error. First, the PROPELLER algorithm depends on choosing an appropriate “reference” blade, to which other blades are aligned. The iteratively found solu- tion following a poor choice of the reference blade may lead to a local minimum in error space that is not the global minimum. A second potential source of error is the fact that simply choosing the shift that gives the maximum correlation between blades gives no regard to the preci- sion of this fit (e.g., from system noise and artifact). This precision is related to the local change in correlation about that peak, as described in the theory section. A new shift estimation algorithm is presented in this work that addresses these areas of potential error. Figure 1 shows the flowchart for the overall PROPELLER recon- struction algorithm and its rotation and translation esti- mation algorithms, along with the original and proposed methods for shift estimation used for rotation and trans- lation estimation. These algorithms are presented in the Theory section, and preliminary tests with volunteers are outlined in the Methods and Results sections. THEORY This section reviews the top-level original motion esti- mation and correction algorithms, as well as the low- level method for calculating correlation between two 1 Barrow Neurological Institute, Phoenix, Arizona, USA. 2 Philips Healthcare, Phoenix, Arizona, USA. *Correspondence to: James Pipe, Ph.D., Keller Center for Imaging Innova- tion, Barrow Neurological Institute, 350 W. Thomas Road, Phoenix, AZ 85013. E-mail: jim.pipe@thebni.org Received 10 April 2013; revised 24 July 2013; accepted 29 July 2013 DOI 10.1002/mrm.24929 Published online 4 September 2013 in Wiley Online Library (wileyonlineli- brary.com). Magnetic Resonance in Medicine 72:430–437 (2014) V C 2013 Wiley Periodicals, Inc. 430