SU‐F‐BRF‐05: Influence of Rigid Registration On the Deformable Registration Quality in the Context of Atlas Contouring

Purpose:Deformable registration requires a rigid registration as an initialization. Normally, a global rigid is assumed acceptable. However, for many use‐cases, such as atlas contouring, there is no obvious “best” global rigid. Thus, we have investigated how a rigid registration affects the deformable registration in the context of atlas contouring and if a “global best” rigid registration can be found.Methods:All possible combinations of source and target images from a database of 7 expertly delineated head and neck images (26 organs) were evaluated. For each source‐target‐pair, the image centroids were aligned, and subsequently rigid registrations were simulated by sweeping the target over an isotropic grid (−100mm to +100mm with respect to centroid alignment, in 10mm steps). Deformable registration was performed following each initialization using commercial software (RTx, Mirada Medical, Oxford, UK). Contours were transformed from source to target. DICE coefficients between the transformed and original contours were calculated on the target. Successful registration was taken to be where DICE was within 80% of the maximum achievable. The capture range was determined by fitting a least‐squares ellipsoid to the space of initialisations resulting in success. The mean of the ellipsoid centers determined the “best” global rigid registration, and the distance of each ellipsoid center to this was computed.Results:The extent of the capture range averaged over all organs and cases was x=64.3mm, y=51.4mm, z=32.3mm. The standard deviation of the ellipsoid centers from the “best” rigid was x=37.8mm, y=56.6mm, z=38.6mm. This per‐organ deviation from the global rigid is in the same order of magnitude than the capture range of the deformable algorithm.Conclusion:Finding the “best” rigid registration is an ill‐posed problem, as the optimum rigid registration will differ between target organs. Thus, having an algorithm with a large capture range is essential to a good deformable registration.All authors were employed by Mirada Medical Ltd., Oxford, Oxfordshire, United Kingdom.