Results of Registration

The training and registration algorithms described above were tested using MR datasets from 37 patients¹. Each patient was scanned using two protocols, a coronal SPGR scan of $256 \times 256 \times 124$ voxels with a voxel size of $0.9375 \times 0.9375 \times 1.5$ mm and a dual echo (proton density and T2-weighted) axial scan of $256 \times 256 \times 52$ voxels with a voxel size of $0.9753 \times 0.9375 \times 3.0$ mm. One of the patients' datasets was used for training purposes and the remaining 36 were used for testing. The two types of registration performed in these experiments were PD with T2 and SPGR with PD. The joint intensity distribution for each modality pair was modeled using a Parzen estimation. The initial pose in all cases was about $90^\circ$ and a few centimeters away. Each registration converged in 1-2 minutes on a Pentium Pro 200. Table 1 shows the results of the registrations.

 

Table: The results of registering 36 test images. The registration error (in mm) was computed by taking the average error of the eight vertices of the imaging volume.
$^\ast$ The patient moved between scans and thus the ground truth is wrong. By inspection, our alignment looks better aligned than the ``ground-truth'' registration.

PD / T2 Registration

SPGR / PD Registration

#	Error
1	0.74
2	0.89
3	0.86
4	0.79
5	0.90
6	0.80
7	0.81
8	0.80
9	0.82

#	Error
10	0.79
11	0.73
12	0.68
13	1.52
14	0.63
15	0.80
16	0.89
17	0.82
18	1.36

#	Error
19	1.42
20	0.75
21	0.69
22	1.33
23	0.66
24	1.01
25	1.07
26	0.70
27	0.81

#	Error
28	1.26
29	0.76
30	2.24
31	0.90
32	0.78
33	1.28
34	0.74
35	1.08
36	0.89

#	Error
1	4.87*
2	2.69
3	0.16
4	0.78
5	1.69
6	5.59*
7	3.02*
8	1.36
9	0.73

#	Error
10	1.47
11	3.17
12	1.32
13	1.18
14	1.59
15	1.25
16	1.26
17	1.19
18	1.85

#	Error
19	2.57
20	2.67
21	2.77
22	3.47
23	3.87
24	6.03*
25	9.72*
26	2.03
27	5.03*

#	Error
28	1.54
29	2.15
30	1.50
31	1.43
32	1.83
33	1.64
34	3.18
35	2.01
36	2.46