Room: Osceola Ballroom C
Purpose: To evaluate the deformable image registration (DIR) performance of Velocity, MIM, and Eclipse, and to characterize inter-algorithm and inter-user variability for these algorithms using data submitted by multiple institutions.
Methods: 35 institutions performed DIR on the Deformable Image Registration and Evaluation Project (DIREP) virtual head and neck phantoms, using either Velocity, MIM, or Eclipse. The resulting deformation vector fields (DVFs) were compared to the ground-truth DVFs, and the target registration error (TRE) was calculated for six ROIs: external, brainstem, cord, mandible, left parotid, and right parotid. Then, statistical analysis was performed in order to determine the algorithmic and user dependence on DIR performance.
Results: Overall registration accuracy was found to be 2.04 Â± 0.35 mm for Velocity, 1.10 Â± 0.29 mm for MIM, and 2.35 Â± 0.15 mm for Eclipse. For this data, MIM produced results with an average TRE up to 1 mm smaller than either Velocity or Eclipse. Both Velocity and Eclipse produced similar results. User dependence within an algorithm was not found in most cases. When user dependence was identified, its impact on DIR performance was much smaller than the variability introduced by choice of algorithm.
Conclusion: Choice of DIR algorithm appears to have a larger impact than does the user dependence within a given algorithm. DIR commissioning should occur after software updates, as significant changes in DIR accuracy can be introduced. Because this can be time consuming, tools provided by vendors to expedite the commissioning process in accordance with TG-132 recommendations would be beneficial. However, the tolerance for maximum registration error recommended by TG-132 may not be achievable with current DIR software in some cases.