SU-F-J-134: Evaluation of Deformation Field Accuracy Using Regional Rigid Registrations with Variable Size.

PURPOSE: How to quickly assess the uncertainty in the Deformation Field (DF) generated from Deformable Image Registration (DIR) over the imaged anatomy is a major issue. Here we introduce a method to determine the DF uncertainty using local rigid registrations of varying box size.

METHODS: A series of 6-degree rigid registrations are performed with a cubic ROI centered on the voxel of the interest. The size of the registration box is varied from a large value, e.g. 30mm to 3mm. The translation vector obtained from the registration of the largest box is assumed to be most reliable and the reliability of the registrations with smaller regions is assumed to be good as long as the variation in the translation vector, as the box size is reduced, is less than the reduction amount in the box size at each step. In this way a smallest reliable box (SRB) size is determined. A distance uncertainty (DU) is calculated for each voxel based on (1) half size of SRB and (2) the magnitude difference between the SRB translation vector to the voxel's DF. This DU is compared to other DIR reliability measures such as Jacobian determinant and root mean square.

RESULTS: The DU for selected voxels was found to agree with those expected from various organ types. For example, mean DU from the voxels in rectum and bladder was 8.4 and 8.9mm respectively, while the DU of the mean femoral head was 2.2mm. 88% of voxels with small DU (<3mm) resulted in reliable Jacobian determinant range (0.5-2), while DU>5mm voxels had 69% of voxels in this range.

CONCLUSION: A method to assess DIR reliability is introduced. The method's ability to provide uncertainty information in terms of distance could be an advantage over other measures such as Jacobian or intensity difference.