is a vast subject, and the purpose of the work described in this thesis was to begin to address the issue of accuracy.
Some of the aspects of this topic are discussed here but a large number of questions are still unanswered. The contributions of this work on the understanding of the accuracy problems in and suggestions for further work in this area, are discussed below.
In Chapter 2 a general model of was introduced. This model presented the image acquisition, the digitizing of the patient's head and the registration as being the links between various spaces involved in , and clearly showed how the various elements of interact with each other to produce the overall result.
It was shown in Chapter 3 that the calculation of the RMS registration error over a large number of registrations can be a useful technique to characterize the performance of different registration methods or different conditions of application of the same registration method. Moreover this method has the advantage of characterizing the variation, over the volume of the head, of the registration error.
The comparison between homologous point matching and surface matching registration techniques did not give the accuracy of each in absolute terms but allowed a good impression to be obtained of the order of magnitude of the registration errors that can occur in a real clinical situation. Moreover, it was demonstrated that a small rotational error can cause the registration error to rapidly increase with the distance from the center of the head. The large differences observed between homologous point and surface matching registrations can in part be explained by the reasons below.
Simulations described in Chapter 3 showed that a sub millimetric accuracy over the whole field of view can be achieved, in noisy situations, by using a large number of homologous points (see, for example, simul_50_2simul_50_3). Since it is difficult in practice to find more than half of a dozen landmarks well defined on both the patient and the image, the only way to identify a larger number of homologous points is to use extrinsic landmarks. An MR distortion experiment (Chapter 5) showed however that the use of such extrinsic structures for registration must be approached with care. Susceptibility effects can introduce shifts that may substantially reduce the registration accuracy (but does not affect the precision, i.e., the variability from one registration to the other). Moreover, the fact that the registration error decreases as the square root of the number of homologous point pairs used implicitly assumes that the error on the localization of the points is random. However, distortion due to susceptibility effects can introduce non-random errors that would make the use of a large number of fiducial landmarks ineffective in reducing the registration error.
The use of the three-point Dixon technique for the acquisition of field maps has been illustrated and it has been shown to be potentially useful in the determination of patient-induced field inhomogeneities. Its efficient use requires however that a special pulse sequence be implemented, which is not the case at the MNI. Also, more work needs to be performed in order to obtain a better phase unwrapping over the region of the skin.
The accuracy of the localizing device (the FARO surgical arm) used at
MNI, was assessed by a simple experiment using a mechanical ``gold
standard'' described in Chapter 4. In addition, the
accuracy of another localizing device, the OPTOTRAK, was measured by
the same technique. The result have shown that the OPTOTRAK gave
distances both closer to the true values and with less variability,
with its average accuracy having been evaluated as 
0.09 mm with
a maximum error on the distance measured of 0.18 mm. The average
accuracy of the FARO arm was evaluated as 
0.25 mm, with a
maximum error of 0.57 mm, which is acceptable considering the present
level of accuracy observed in .