Hi,
My question relates to the order of processing images from different time points in the longitudinal stream.
The base-line images I have were collected in 2006, but many of these images are of poorer quality than those collected in 2008 at follow up.
Given this description from Han et al., 2006:
"The longitudinal scheme differs from the original procedure in three major steps: preprocessing, intensity normalization and surface deformation. *Assuming that a series of scans of the same subject are obtained at different time points, the data from time point one is first processed using the original procedure. For the processing of later time points, the three major steps mentioned above are modified.* First, a linear registration is computed between the image volume of a later time point and that of the first time point. Note that the volume itself is not transformed, but only the registration matrix is stored and will be used in later steps. Second, at the intensity normalization step, instead of re-computing the WM control points, the control points (automatically computed) from time point one are mapped to the current volume using the previously computed linear registration, which are then used to estimate the bias field. Third, at the surface deformation step, instead of starting the deformation from the topology-corrected WM tessellation, the gray/white and pial surfaces reconstructed from time point one are first transformed using the linear registration and then used to initialize the deformation for the gray/white and pial surfaces respectively for the current time point. Such an initialization scheme reduces the problem of deformable model methods being sensitive to initialization and local optimality (Dale et al., 1999; Xu et al., 1999; Han et al., 2005a). It also eliminates variation in initial surface tessellation and topology correction since these two steps are no longer needed in the processing of later time points."
Would it make sense to use the better quality images, which in this case are those collected at time point 2, as the "baseline" when processing in the longitudinal stream to get a more accurate result for time point 1?
Additionally, is there any consensus about the degree to which the longitudinal processing stream assists in overcoming variance due to multi-site scanning?
Thanks in advance for any help on these issues,
Kind Regards,
Meg Dennison ------------------------------- PhD Candidate / Masters Clinical Psychology Psychological Sciences Melbourne University -------------------------------
Hi Meg,
the new longitudinal stream creates a base/template from all time points (it is not using the first one specifically). Still if the first is very different from the rest, the quality of this template will be influenced. If you do not want to remove the first one, just try it out and see what happens (using FS 4.5). Also there is a documentation on the wiki. Han's description below is for the old stream that was biased with respect to the first time point.
The longitudinal stream significantly reduces variability. This is also true for multi-site scanning, different software versions on the scanner or different scanners. Although all of these are not very desirable scenarios.
Best, Martin
On Thu, 2010-06-03 at 15:11 +1000, Meg Dennison wrote:
Hi,
My question relates to the order of processing images from different time points in the longitudinal stream.
The base-line images I have were collected in 2006, but many of these images are of poorer quality than those collected in 2008 at follow up.
Given this description from Han et al., 2006:
"The longitudinal scheme differs from the original procedure in three major steps: preprocessing, intensity normalization and surface deformation. Assuming that a series of scans of the same subject are obtained at different time points, the data from time point one is first processed using the original procedure. For the processing of later time points, the three major steps mentioned above are modified. First, a linear registration is computed between the image volume of a later time point and that of the first time point. Note that the volume itself is not transformed, but only the registration matrix is stored and will be used in later steps. Second, at the intensity normalization step, instead of re-computing the WM control points, the control points (automatically computed) from time point one are mapped to the current volume using the previously computed linear registration, which are then used to estimate the bias field. Third, at the surface deformation step, instead of starting the deformation from the topology-corrected WM tessellation, the gray/white and pial surfaces reconstructed from time point one are first transformed using the linear registration and then used to initialize the deformation for the gray/white and pial surfaces respectively for the current time point. Such an initialization scheme reduces the problem of deformable model methods being sensitive to initialization and local optimality (Dale et al., 1999; Xu et al., 1999; Han et al., 2005a). It also eliminates variation in initial surface tessellation and topology correction since these two steps are no longer needed in the processing of later time points."
Would it make sense to use the better quality images, which in this case are those collected at time point 2, as the "baseline" when processing in the longitudinal stream to get a more accurate result for time point 1?
Additionally, is there any consensus about the degree to which the longitudinal processing stream assists in overcoming variance due to multi-site scanning?
Thanks in advance for any help on these issues,
Kind Regards,
Meg Dennison
PhD Candidate / Masters Clinical Psychology Psychological Sciences Melbourne University
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu
-
Martin Reuter -
Meg Dennison