Intra and Intermodality Validation of Registration Algorithms On a Deformable Phantom
S D.Vincent*, L Archambault, CHUQ Pavillon Hotel-Dieu de Quebec, Quebec, QCSU-E-J-70 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
Purpose: Clinical use of deformable image registration algorithms (DRA) requires a thorough preliminary study of their capacity and performance. To rigorously quantify the registrations, images showing deformation of precise amplitude must be used. Deformations representative of the average anatomical variations possibly encountered during radiation treatments were used for DRA evaluation.
Methods:A three-dimensional anthropomorphic phantom representing the main pelvic structures (prostate, bladder, rectum, femoral heads) was created. Three types of deformation were studied: translation (prostate), volume variation (bladder) and deformation of contours without volume change (prostate). The design of the phantom allows a precise knowledge of deformation amplitudes. Each pairs of initial and deformed images were scanned on both CT and CBCT. This way an image database offering the same images on both modalities was available for intra and intermodality validation. The algorithm validated was an open-source software (Elastix) that offers a large control of the registration parameters.
Results:Deformation of 5 mm on the contour, translation of 0.5, 1, 2 cm and volume variation of 50, 100, 200 ml were scanned on both modalities. Registration with basic Mutual information metric succeed between 98% and 71%, these results representing the fraction of the deformation registered with success on its initial image. The worst cases were observed when a CT is registered onto a CBCT. By combining the Mutual information registration metric with manually delineated points (corresponding point metric) the worst cases (71 to 79%) were greaty improved (88 to 96%).
Conclusion:By optimizing the deformation parameters in a controlled setting (i.e. an anthropomorphic phantom) we could achieve better DRA efficiency. Combining multiple registration metric lead to registration of superior quality. Deformable vector field shows that many areas free from actual deformation were altered during registration which indicate that deformation parameters optimized to improve registration of organ contours are not well suited for dose deformation.