Impact of slice selection for the EPI 3D reference and isotropic resampling on the estimation of motion parameters

[Gab-D-G]

With the original strategy, the first 50 frames are selected to generate the reference target for estimating motion parameters. When there is a drift in motion across time, the distance from the reference image is minimal at the beginning (since the reference is from the initial frames), and increases over time. This can impact in turn estimates of framewise displacement (FD), where the FD amplitude tends to increases over time (see example below, and related post #259 ):

In this example, the scan had minimal motion during acquisition, with the exception of a slow drift in the brain's position across time. The 6 motion parameters accurately model the brain drift, but we can also observe that the baseline amplitude of FD estimate is increasing over time. This trend in FD estimates is likely artefactual, and could introduce false differences in motion estimates across different timepoints in the EPI acquisition. The artefact could arise for instance from partial voluming effects as the brain shifts from its original position, changing the brain contrast/orientation across time, deviating progressively further away from the reference EPI generated from the initial frames, and in turn lead to increased registration instability (i.e. more noise in the motion estimates, which will increase baseline FD measures).

To address this issue, we explore two solutions: uniform frame sampling across time to generate the reference image (i.e. take 50 frames uniformaly distributed across the entire timecourse of acquisition), and isotropic resampling of the EPI (i.e. resample the EPI to isotropic resolution prior to estimation of motion parameters). The uniform sampling of time frames would avoid generating a reference with brain contrast biased at a particular point in time, and the isotropic resampling should mitigate registration instabilities arising from partial volume effects (i.e. reduce noise in motion estimates).

Below is the same scan after using a reference with uniform sampling of frames across time:

We can see that the drift is now centered in the middle of the timecourse, which would be expected from sampling across the entire timecourse, but importantly, there is less evidence of a trend in FD, and the peak amplitude in FD was almost halved (from around 0.04mm to around 0.02mm). There is still however noticeable noise in the 6 motion parameters, which seem to be increasing as the brain is drifting from the center of reference.

Below is the same estimates after using the uniform reference image, and resampling the EPI to isotropic resolution prior to motion estimation. The isotropic resampling consisted of converting the image from 0.25x0.25x0.5mm to 0.25mm isotropic resolution, thus reducing the slice thickness along the sagittal axis.

In this case, the baseline noise of the motion estimate has been greatly reduced, and peak amplitude in FD is close to 0.01.

These can be important considerations if the data analysis is sensitive to timing at acquisition. Further software updates will integrate the two features to mitigate these issues.