The registration algorithm we have previously developed [Reference and demo] estimates the transformation between matched frames and its uncertainty. To give a more intuitive information about the registration accuracy, we compute the typical boundary accuracy (the mean standard deviation on image corners due to the registration uncertainty) and the typical object accuracy (the mean standard deviation on matched extremal points due to the registration uncertainty only).
In the CT matching experiment, images are 160 x 200 x 128 and the voxel size is 1.3 x 1.3 x 2 mm. We obtain the following accuracy for 550 frame matches:
|
|
|
|
|
|
In the MRI matching experiment, images are 256 x 256 x 54 with a voxel size of 1 x 1 x 3 mm. For 350 frame matches, we obtain an accuracy of:
|
|
|
In both cases, the number of matches and the accuracy of frames are sufficient to obtain a registration accuracy below the tenth of a voxel.
[ Back to substructure
matching with frames ] 
