Program Information
Quality Assurance of Deformable Image Registration in Radiotherapy
W.T. Watkins1, K. Bzdusek2 , and J.V. Siebers1 ,(1)University of Virginia Health System, Charlottesville, VA (2) Philips Healthcare, Fitchburg, WI
Presentations
SU-F-BRF-3 Sunday 4:00PM - 6:00PM Room: Ballroom FPurpose:To introduce methods to analyze Deformable Image Registration (DIR) and identify regions of potential DIR errors.
Methods:DIR Deformable Vector Fields (DVFs) quantifying patient anatomic changes were evaluated using the Jacobian determinant and the magnitude of DVF curl as functions of tissue density and tissue type. These quantities represent local relative deformation and rotation, respectively. Large values in dense tissues can potentially identify non-physical DVF errors. For multiple DVFs per patient, histograms and visualization of DVF differences were also considered. To demonstrate the capabilities of methods, we computed multiple DVFs for each of five Head and Neck (H&N) patients (P1-P5) via a Fast-symmetric Demons (FSD) algorithm and via a Diffeomorphic Demons (DFD) algorithm, and show the potential to identify DVF errors.
Results: Quantitative comparisons of the FSD and DFD registrations revealed <0.3 cm DVF differences in >99% of all voxels for P1, >96% for P2, and >90% of voxels for P3. While the FSD and DFD registrations were very similar for these patients, the Jacobian determinant was >50% in 9-15% of soft tissue and in 3-17% of bony tissue in each of these cases. The volumes of large soft tissue deformation were consistent for all five patients using the FSD algorithm (mean 15%±4% volume), whereas DFD reduced regions of large deformation by 10% volume (785 cm³) for P4 and by 14% volume (1775 cm³) for P5. The DFD registrations resulted in fewer regions of large DVF-curl; 50% rotations in FSD registrations averaged 209±136 cm³ in soft tissue and 10±11 cm³ in bony tissue, but using DFD these values were reduced to 42±53 cm³ and 1.1±1.5 cm³, respectively.
Conclusion: Analysis of Jacobian determinant and curl as functions of tissue density can identify regions of potential DVF errors by identifying non-physical deformations and rotations.
Funding Support, Disclosures, and Conflict of Interest: Collaboration with Phillips Healthcare, as indicated in authorship.
Contact Email: