Longitudinal data

The procedure is in general the same as described for cross-sectional data. The main difference is that all images of one subject will be registered to correct for position (but not size) and the normalization estimates are derived from the first (baseline) scan only. The estimated normalization parameters are then applied to all images of one subject. The idea of this procedure is that spatial normalization should not remove differences between the scans of one subject. A previous version of this longitudinal processing algorithm was applied in two key papers investigating structural changes induced by learning (Draganski et al. 2004/2006).

Additional bias correction between time points

Because longitudinal scans are often acquired with a large gap between the time points this may cause different distributions of intensity nonuni-formities. Although each scan will be corrected independently using bias correction subtle but systematical differences in the bias field remain and may influence results of segmentation. To minimize these intensity differences between the time points we approximate the difference bias field by using the intracraniel parts of the difference image which is smoothed with a very large gaussian kernel of around 30 mm. However, more exact methods might exist for this purpose.


bias-longitudinal
This figure shows an example of inhomogeneities between two longitudinal scans. The time gap between the scans was 1 year and after subtraction of the normalized T1 images the inhomogeneities between these two T1 images are obvious in the upper left image and also in the segmented images in the upper right side. We approximate these spatial low-frequency effects by smoothing the difference between the T1 images and restrict smoothing to intra-cranial parts only. After removing these inhomogeneities the differences between the T1 images are locally minimized (lower left image) and the difference between the segmented images is also corrected (lower right image).