To evaluate non-linear image registration for distortion correction in human optic nerve diffusion imaging using edge plots as image alignment metrics.Among the many challenges of high resolution in vivo human optic nerve diffusion imaging is the effect of eye-ball movement, which can lead to substantial intraorbital optic nerve motion. Such optic nerve motion is non-linear and independent of head movement. These non-linear optic nerve movements further couple with the distortions in diffusion echo planar imaging (EPI) due to field inhomogeneity and eddy currents (Fig. 1). Existing distortion correction schemes routinely used for brain diffusion MRI fail to adequately correct for these non-linear distortions specific to optic nerve diffusion imaging. In this study, we developed and evaluated a non-linear registration scheme to address this issue.Oblique axial optic nerve diffusion MRI data (Fig. 2) were acquired in six adult subjects on a 3T scanner (Trio, Siemens) with a 32ch head coil (12 anterior Rx elements only, Siemens). An inner volume imaging spin echo EPI diffusion sequence was used to acquire reduced FOV optic nerve images at 1.3 mm isotropic resolution [1], TR/TE=5000/56.4 ms, FOV=166 x 41.5 mm, matrix=128 x 32, 6/8 partial Fourier, ETL=24 (20.6 ms), monopolar diffusion encoding, optimized 25 multi-b_val multi-b_vec diffusion scheme with b_max=1000 s/mm2, T_acq=2.5 min. x 4 effective averages.All frames in the diffusion acquisition were corrected for distortion and motion with FSL/topup [2] to estimate field maps from a pair of phase-encoding reversed b0 images and further with FSL/eddy [3] using quadratic eddy current term estimations.To correct for residue image distortion after FSL/eddy, a non-linear registration scheme (Fig. 3) adapted from a motion correction scheme for spinal cord diffusion imaging [4] was developed. (i) All frames were split into b0 frames (b<90 s/mm²) and diffusion weighted (DW) segments (D_i); each segment consists of temporally adjacent DW frames between two adjacent b0 frames. (ii) All b0 frames were registered to the first b0 image (b0-f₀). (iii) The DW frames in each Di were registered to the first DW frame of the segment (D_i-f₀). (iv) All DW segments were registered to the first DW segment (D₀-f₀). (v) All DW segments were registered to the first b0 frame. Linear registration (FSL/flirt with correlation ratio as cost function) was applied in step (v) because of the different contrast between DW and b0 frames. Non-linear registration was applied in steps (ii-iv). We evaluated three non-linear registration methods: (1) FSL/fnirt with sum-of-squared difference as cost function, (2) ANTs [4] with neighborhood cross correlation (cc) as cost function, and (3) ANTs with mean squared intensity differences (ms) as cost function.The optic nerve registration results were evaluated by comparing unprocessed, FSL/eddy unwarped, FSL/eddy+FSL/fnirt, FSL/eddy+ANTs/cc, and FSL/eddy+ANTs/ms registered images. We defined the two optic nerve edges on a horizontal or vertical line from a coronal slice as the edges between two adjacent voxels whose intensity difference are the maximum or minimum among all the edges on the line. The frequency of an edge between two adjacent voxels along the specified line was defined as the number of frames with estimated optic nerve edge fell on that edge location. For visualization, the optic nerve edges along the horizontal line (8 voxels) of a representative coronal slice (Fig. 4C), stacked along all frames, were colored as red / yellow (Fig. 4A). All stacked edge plots (Fig. 4A) were overlaid on top of each other without background image (Fig. 4B) for comparison. Histograms were shown on the top of the plots to indicate the frequencies of the optic nerve edges (Fig. 4A, B). In addition, the percentage of edge frequency was color coded and overlaid on the averaged image of all diffusion measurement frames on two-dimensional sub-region of the representative coronal slice (Fig. 5).

In all six subjects, FSL/eddy plus the additional non-linear registration scheme outperforms FSL/eddy unwarping alone, which yields consistently better results than unprocessed images. We observed a trend towards better non-linear registration results using FSL/eddy+ANTs/ms or FSL/eddy+ANTs/cc than FSL/eddy+FSL/fnirt, although a conclusion cannot be drawn from these preliminary results.

The additional non-linear registration step following FSL/eddy unwarping further improved the alignment of the optic nerve edges in the diffusion imaging frames. Such optic nerve edge alignment improvement is expected to be crucial for the accurate diffusion parameter estimation in diffusion MRI signal modeling.