SU-E-J-99: Computation of Deformable Image Registration Confidence Maps

Abstract

Purpose: Many metrics used to assess Deformable Image Registration (DIR) quality are overly simplistic while also being time consuming. In this study, we develop an automated method to obtain an estimate of the DIR errors at every point on the registration grid. This Deformable Registration Confidence Map (DRC-Map) can then be overlaid onto the deformed volume to highlight regions of poor registration quality. Methods: The DRC-Map computation process for volume A deformed onto volume B consists of four main steps: i) Deform B onto A, ii) Apply a random perturbation to the deformation vector field (DVF) of this deformation and smooth the Result, iii) Apply the perturbed DVF to volume B to create an approximation of A (A’), and iv) Deform A’ back onto B. The perturbed DVF along with the DVF obtained from step 4 can be used to explicitly track the motion of each voxel is they deform from volume B to A’ then back to B. Three patients, two liver and one pancreatic, were chosen for DRC-Map computation. For each patient, volume B was defined by applying a known DVF to volume A. A 5% random perturbation was used to create A’. The DRC-Map was compared to the true errors obtained through knowledge of the exact transform from A to B given by the pre-defined DVF. Results: For the patients considered, the mean difference between the estimated and true registration errors were 1.15 mm, 1.14 mm, and 1.08 mm; with 80% of the DRC-Map values within 1.52 mm, 1.49 mm, and 1.50 mm of the true errors respectively. Conclusion: Computation of a DRC-Map can be fully automated and does not require organ delineation or landmark identification. DRC-Maps provide a fast and accurate mechanism for highlighting regions of poor DIR results in a spatially dependent manner.