preview

Mean And Standard Deviance Of Different Methods In Intra-Subject Experiment

Decent Essays
Open Document

Table [tab:test1_rmde], [tab:test1_rmie], [tab:test1_nnjd] and [tab:test1_time] respectively show the mean and standard deviation of the 6 compared methods on 5 datasets using the RMDE, RMIE, NNJD, and the running time metrics.

Mean and standard deviation of RMDE of 6 methods on 5 datasets in the intra-subject experiment. The second last row shows the p-value of a paired two-tailed t-test between the pwBsp method and the method listed in each column. The last row shows the difference of the RMDE mean (\mu-diff) of all image pairs between pwBsp and the other method in each column.

Mean and standard deviation of RMIE of 6 methods on 5 datasets in the intra-subject experiment. The second last row shows the p-value of a paired two-tailed …show more content…

In Table [tab:test1_nnjd] DRAMMS was the only method generating deformation fields with non-positive Jacobian determinant. Though the number of violations was considered small when compared to the total voxel number in the image. The ANTs method had none violation because it was built diffeomorphic. Thanks to the proposed regularization control strategy, especially the Jacobian determinant based filtering algorithm, the globalDct, pwAffine and pwBsp had zero violations.

Among the compared methods, the Elastix method ran the fastest consistently, as shown in Table [tab:test1_time]. The Elastix's speed advantage can be attributed to the ITK platform, in which implementation was extensively optimized. Though the ANTs method was also implemented based on the ITK platform, its speed was probably slowed down by the symmetric diffeomorphic framework. The diffeomorphic feature required a constant velocity field. The DRAMMS increased the feature space thus having a higher computational cost. In addition, it was not implemented with any multi-processing technique.

In the intra-subject experiment, we can conclude that the proposed approach demonstrated improved performance when compared to the state-of-the-art methods. Even though the proposed approach was not the most efficient one, it provided a good balance between accuracy and computational cost.

The registration

Get Access
Related
Get Access