Biomedical Engineering Reference
In-Depth Information
In the case of retrospective multimodality registration, the correct registra-
tion parameters are typically unknown. Various evaluation methods have been
used to assess registration accuracy, including phantom validation, observer as-
sessment [11], fiducial marks [8, 9, 10], and cross validation [47]. Since the Tx
and Em scans used in our study are intrinsically registered (see [43, 44, 45, 46]),
we have us an ideal method to evaluate the registration accuracy. In this special
case, the accuracy of a registration technique can be assessed by evaluating the
registration parameters that are generated when these Tx and Em images are
used as input.
The Tx and Em images were also individually registered to the MR images.
Based on those registration results, the registration parameters between the
Tx and Em images were indirectly calculated. If we denote the transformation
matrix relating the MR and Tx images,
M
MR
/
Tx
, and the matrix relating the MR
and Em images,
M
MR
/
Em
, the matrix relating the Tx and Em images can be
calculated as
M
Tx
/
Em
=
M
−
1
MR
/
Tx
M
MR
/
Em
.
For this indirect registration, two sets of experiments were performed. In
the first set, no attempt was made to manually preregister the images before
the automatic registration started. In the second set, all images were initially
brought into approximate registration using a manual method. The same manual
adjustments were used in cross-entropy, reversed cross-entropy, and symmetric
divergence registrations.
If the MR/Tx and MR/Em registrations deviate from the true registration
in the same manner, the indirect registration evaluation may yield a mislead-
ing registration accuracy. To guard against this, all registration results were
visually checked. Only if both registrations were visually acceptable, were
their registration results used to indirectly calculate the Tx/Em registration.
Since it is reported in Studholme
et al.
[11] that clinicians can detect registra-
tion parameter differences of 4
◦
in
x
and
y
rotation angles, 2
◦
in the
z
rota-
tion angle, 2 mm in
x
and
y
translations, and 3 mm in
z
translation, the reg-
istrations were believed to be close to the truth if the results were visually
acceptable.
The accuracy and robustness of direct and indirect Tx/Em registrations were
analyzed in terms of the mean and standard deviation of the differences between
the resultant registration parameters and the intrinsic ones. Alternatively, one
can compute the difference of a group of selected points. As reported in West
et al.
[8], the mean of point differences is related to the rotation and translation
