Table 1.

Test result for atlas-to-image registration.

	Deformation Error for each voxel [mm]							detJ > 0
Data percentile for all voxels	0.3%	5%	25%	50%	75%	95%	99.7%
Affine	0.0613	0.2520	0.6896	1.1911	1.8743	3.1413	5.3661	N/A
D, velocity, stride 5	0.0237	0.0709	0.1601	0.2626	0.4117	0.7336	1.5166	100%
D, velocity, stride 14	0.0254	0.075	0.1675	0.2703	0.415	0.743	1.5598	100%
D, deformation, stride 5	0.0223	0.0665	0.1549	0.2614	0.4119	0.7388	1.5845	56%
D, deformation, stride 14	0.0242	0.0721	0.1671	0.2772	0.4337	0.7932	1.6805	0%
P, momentum, stride 14, 50 samples	0.0166	0.0479	0.1054	0.1678	0.2546	0.4537	1.1049	100%
D, momentum, stride 5	0.0129	0.0376	0.0884	0.1534	0.2506	0.4716	1.1095	100%
D, momentum, stride 14	0.013	0.0372	0.0834	0.1359	0.2112	0.3902	0.9433	100%
D, momentum, stride 14, 40 epochs	0.0119	0.0351	0.0793	0.1309	0.2070	0.3924	0.9542	100%
D, momentum, stride 14 + correction	0.0104	0.0309	0.0704	0.1167	0.185	0.3478	0.841	100%

The table shows the distribution of the 2-norm of the deformation error of the predicted deformation with respect to the deformation obtained by numerical optimization. Percentiles of the displacement errors are shown to provide a complete picture of the error distribution over just reporting the mean or median errors over all voxels within the brain mask in the dataset. D: deterministic network; P: probabilistic network; stride: stride length of the sliding window for whole image prediction; velocity: predicting initial velocity; deformation: predicting the deformation field; momentum: predicting the initial momentum; correction: using the correction network. The detJ > 0 column shows the ratio of test cases with only positive-definite determinants of the Jacobian of the deformation map to the overall number of registrations (100% indicates that all registration results were diffeomorphic). Our initial momentum networks are highlighted in bold. The best results are also highlighted in bold.