Comparison with Existing Methods

A comparison of FLIRT with several other registration packages was initially performed using the consistency test explained above. The other registration packages used were AIR [19], SPM [6], UMDS [16] and MRITOTAL [3]. These methods were chosen because the authors' implementations were available, and so this constituted a fair test as opposed to a re-implementation of a method described in a paper, where often the lack of precise implementation details makes it difficult to produce a good working method. The particular experiment that was performed was inter-subject and inter-modal using 18 different images as the floating images (like the one shown in Figure 9), all with the MNI 305 brain [3] as the reference image. The 18 images were all $256 \times 256 \times 30$, T2-weighted MR images with voxel dimensions of 0.93mm by 0.93 mm by 5mm, while the MNI 305 template is a $172 \times 220 \times 156$, T1-weighted MR image with voxel dimensions of 1mm by 1 mm by 1mm.

Figure 9: Example slices from one of the images used in the consistency study (after registration). The red lines represent edges from the standard image (the reference image) overlayed on the transformed initial image.

The results of one such test, using six different rotations about the Anterior-Posterior axis, are shown in Figure 10. It can be seen that only FLIRT and MRITOTAL performed consistently. This indicates that the other methods (AIR, SPM and UMDS) frequently get trapped in local minima, i.e. are not as robust. A further consistency test was then performed comparing only MRITOTAL and FLIRT. This test used initial scalings rather than rotations. The reason that this is important is that MRITOTAL uses a multi-resolution local optimisation method (Gradient Descent) but relies on initial pre-processing to provide a good starting position. This pre-processing is done by finding the principle axes of both images and initially aligning them. Consequently the initial alignment compensates for rotations but does not give any information, and hence correction, for scalings. The results of the scaling consistency test are shown in Figure 11. It can be seen that, although generally consistent, in three cases MRITOTAL produces registrations that deviate by more than 20mm (RMS) from each other. In contrast, FLIRT was consistent (less than 2mm RMS) in all cases.

Figure 10: Results of the consistency study, plotting RMS deviation (in mm) versus image number for (a) AIR, (b) SPM, (c) UMDS, (d) MRITOTAL and (e) FLIRT. For each of the 18 source images (T2-weighted MRI images with voxel dimensions of 0.93mm by 0.93 mm by 5mm) there are 6 results corresponding to initial starting rotations of -10,-2,-0.5,0.5,2, and 10 degrees about the $y$-axis (anterior-posterior axis). All of the methods, except FLIRT and MRITOTAL, show large deviations and are therefore inconsistent and non-robust.

Figure 11: Results of the scale consistency study, plotting RMS deviation (in mm) versus image number for (a) MRITOTAL and (b) FLIRT. For each of the 18 source images (T2-weighted MRI images with voxel dimensions of 0.93mm by 0.93 mm by 5mm) there are 6 results corresponding to initial scalings of 0.7, 0.8, 0.9, 1.1, 1.2 and 1.3 about the Centre of Mass. In three cases MRITOTAL shows large deviations and so is less consistent and robust than FLIRT in this case.