A more interesting and more challenging alignment problem consists of registering each patient's coronal SPGR scan with the same patient's PD image. This registration problem is more challenging given that the two modalities were not acquired simultaneously and also do not contain the same region of the head: the PD/T2 images are cropped at the chin (see Figure 1).
Since each patient had the scans performed during the same sitting in
the scanner, the headers of the scans provide the ground-truth
alignment between the various acquisitions, assuming the patient did
not move between acquisitions. However, since the patients' heads
were not fixed in the scanner, patients could move between
acquisitions. Despite this issue, we use the scanner poses as ground
truth, since in most cases it seems that the patient did not move
significantly. By visually inspecting the SPGR registered with the PD
images using the ground truth, one notices a discrepancy of as much as
a centimeter in six of the patients' scans (marked with an
in
Table 1). In these cases, the error values reported
are not valid, and our registration qualitatively appears better
aligned than the ``ground-truth'' registration. Of the 36 cases we
have used in our initial tests of this method, almost all of the cases
automatically registered to within one voxel, from a starting position
of about 90
and a few centimeters away.