2 Longitudinal registration module
1.5) Longitudinal registration 6 0015_LongitudinalRegistration.status
This is a recent add-on to ExploreASL, and this part only runs if there are multiple time points/volumes (multiple T1 scans per subject). This part can be skipped for data without multiple structural scans (TimePoints) per subject. By the loopDB, these multiple T1 scans from the same subject are classically viewed as multiple subjects, since the data structure of ExploreASL is set up for multiple ASL sessions with a single T1, which is valid if we don't expect anatomical brain changes between the ASL scans (as in a CO2 or therapeutic challenge on short timeframe of hours or days). So if anatomical changes can be expected, as in any follow-up study, especially in early-life developmental or late-life atrophying stages, we should account for the longitudinal anatomical changes.
SPM12 longitudinal registration is an extended version of the VBM longitudinal registration. This performs a similar non-linear registration as DARTEL, but instead of assuming that the brains can be completely different (as we expect them to be between subjects) we expect that both T1 images contain the same brain except for shrinkage or expansion. This is a good way to model atrophy/cortical thinning, but also a more realistic way to non-linearly register multiple T1 scans of the same subject, and use DARTEL only to non-linearly register between subjects. Therefore, if longitudinal registration is requested, DARTEL will run only for first volumes/time points of each subjects, otherwise, DARTEL will run for all volumes.
For each TimePoint/volume (i.e. each T1.nii) a LongReg_y_T1.nii file will be created that creates the deformation from the respective T1 image to the average T1 image of the respective subject. For each voxel, this deformation image gives the new coordinates. This deformation image will later be combined with other transformations to transform all images to standard space. Currently, only TimePoints _1 to _9 are allowed (TimePoints>9 needs reprogramming). Same as with the segmentation (see below), by default the biasfield regularization is turned off (resulting in more extensive biasfield modeling) for GE T1 images, since these scanners frequently have a larger bore and consequently a larger biasfield.
When Longitudinal Registration is enabled, the deformation field from native space to MNI space is used from the first TimePoint only, and the LongReg deformation fields are used to warp other TimePoints to the first time point, with these deformations combined for a single interpolation. The first TimePoint can be seen as the one with highest quality, since brain parenchyma and hence segmentation precision are expected to decrease with time. This is why the deformation field to MNI of the first TimePoint is used, and if DARTEL is run as well, this is performed with the first TimePoints only.
It is important to note that segmentation of each individual timepoint/volume will give best results in terms of segmentation/registration, better than longitudinal registration, especially in anatomic anomalies. Therefore, we still segment follow-up time points separately, but don’t use their flow-fields to transform to the common space. Instead, we use the longitudinal registration flow fields and the CAT12 flow fields to transform the segmentations and other data to the common space.
It is important that 1) at the import stage, the volumes are used in subject names as suffix (e.g. SubjectName_1, SubjectName_2 : SubjectName_n);
Examples of the (very subtle) longitudinal registration deformation fields. These look similar to DARTEL flow fields, but with much smaller magnitude of deformations.
Conventionally, Longitudinal registration (LongReg) takes long processing time (~20-45 min) with normal settings. Important settings are age and space (resolution) in which we process our data. When we repeat LongReg twice, with pairwise (2.5 years difference) and serial registration (60 & 62.5 years), the results are identical. So good reproducibility.
Using 1.5x1.5x1.5 mm MNI resolution instead of native space gives a large reduction of computational time, and makes sense from the ExploreASL perspective. Many SPM spatial operations, such as DARTEL run on 1.5x1.5x1.5 mm, resampled in MNI center. It makes sense to use the same quality for longitudinal registration. This space also cuts of the neck and focuses the algorithm on the brain only.
Using different age parameters has an effect on processing time, but this was an effort of the makers to model that with multiple T1ws, more deformations are allowed with longer time intervals than with shorter, as we can expect different magnitudes of atrophy. Hence, we leave this settings as is and use the age of the study participants. If age is not provided as covariate with ExploreASL, 4.5 years is a default optimum trade-off between deformation accuracy and processing time. In cases with wide ranges of expected atrophy, regardless of age, this default should be used as well.
Note: LongReg creates flow fields with a larger FoV. These can be dealt with by the SPM deformation tool, and will resample images into correct space if a y_T1.nii deformation image is provided as well.
This is the QC image for longitudinal registration. Left above = first volume, right above = current volume, below = difference image shown in red over the current volume, left below = before SPM12 LongReg, right is after SPM12 LongReg. When we ignore the different biasfield of both time points (shown as a red hue over the image) we can appreciate how the ventricles have a red lining of their different sizes at the different time points, which is gone after LongReg (right low image) These results are shown without applying the y_T1.nii flowfield, because this would add non-linear deformations to the images, making it harder to focus on the TimePoint differences.