In spinal cord injury patients (SCI), the flow of the spinal cerebrospinal fluid (CSF) has been studied due to its important physiological role. However, the velocity-encoding gradient induces an eddy current which results in a baseline phase offset in MR phase contrast cine flow measurement. A widely adopted method is to use the phase offset in a manually selected adjacent stationary tissue as a reference phase offset.¹ However, the phase offset due to eddy current is spatially inhomogeneous and the reference stationary tissue can include small vessels and motion. Another approach is the image-based extraction of the phase offset.² This method is further developed to study the spinal CSF flow in a large number of participants reliably in an automated procedure.The phase offset image can be segmented into stationary tissue, background, and regions with flow using a spectral decomposition of the cine flow images in a complex data format in the cardiac cycle direction.³ Since the cardiac pulsation and the resulting fluctuation of CSF flow are in a lower temporal frequency, the sum-of-square of the 1st through 3rd harmonic components selects the image regions with fluctuating flow. The phase shift in the flow and background region can be initialized to an averaged phase shift of the stationary tissue. Assuming a very low spatial frequency of the magnetic field induced from the eddy current, the initial composite phase image is smoothed with a 2-D Gaussian smoothing kernel (standard deviation = 1.5). The phase shift at the flow and background is substituted with the smoothed phase shift. The phase shift at the flow and background is updated iteratively by repeating the composition of the phase shift followed by smoothing (Fig. 1). After sufficient iterations, the smoothed composite phase shift can be considered as the estimated phase offset due to the eddy current. The CSF flow velocity in the cranial-caudal direction was measured at the 4th cervical spine level of 10 healthy and 10 chronic SCI participants with an injury below C3 using a 3 Tesla MRI system, the velocity-encoding gradient of 10 cm/s, and a peripheral pulse gating. The cyclic flow rates (a summation of volume flow rates times with the heart period in one cardiac cycle) in the two populations were averaged over the spinal subarachnoid space and statistically analyzed using the Mann-Whitney U-test for the confidence level of p < 0.05.The 0th component represented the stationary tissue, while the higher order components included the vessels, CSF, and tissue with a motion. The CSF flow was mostly collected at the 1st through 3rd harmonic components, while the flow in the carotid arteries was spread out over the spectral range. A flow mask image was constructed by thresholding the sum-of-square image of the 1st through 3rd harmonic components as shown in Fig. 2. The initial phase shift obtained from a smoothed composite phase shift is shown in Fig. 3A. After 11 iterations, the initial phase shift was refined to the phase shift induced by the eddy current as shown in Fig. 3B and Fig. 3C. The phase shift was indeed spatially inhomogeneous with a low spatial frequency. The effect of the phase offset correction was demonstrated as the velocity map at the first cardiac phase (Fig. 4) and the flow curve over two cardiac cycles (Fig. 5). The polarity of velocity was positive for the flow direction from foot to head. Therefore, the negative directional flow corresponded to the systole of the cardiac cycle. It was noticeable that the phase (or velocity) offset varied significantly among subjects in addition to spatial variation within a subject. The cyclic flow rates were in the caudal direction (negative polarity) as -0.057 and -0.014 ml/cycle for the healthy and SCI participants, respectively, with a statistical significance.⁴The phase offset in the phase contrast MR flow images was confirmed to be spatially inhomogeneous and it varied among subjects. The flow region was segmented automatically using a spectral decomposition of the cine flow images. The phase offset on the flow region was estimated using an iterative estimation with a constraint of a low spatial frequency of the phase offset. The proposed phase offset correction method was confirmed by studying the cyclic flow rates of CSF in the spinal subarachnoid space of healthy and SCI participants.