Regularized Deformable Registration of Computed Tomography Images Using Multiple Voxel Resolutions for Four Dimensional Radiation Therapy
M Chao1*, Y Lo1(1) The Mount Sinai Medical Center, New York, NYSU-E-J-67 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
Purpose: Accurate, speedy, and robust image registration is deemed crucial in radiation therapy of cancer management. A multi-resolution deformable image registration (DIR) using regularization strategy was developed for four dimensional radiation therapy.
Methods: A multi-resolution based deformable registration model was implemented with the regularization of energy term to smooth the deformation field. A B-Spline model, although other models such as demons are applicable as well, was employed as the transformation with the Mattes mutual information as the objective function that was to be optimized the limited memory Broyden-Fletcher-Goldfarb-Shanno algorithm (L-BFGS) optimizer. A three level voxel resolution model was utilized for the coarse-to-fine strategy to improve the registration success rate and remove local optima at coarser scales. A regularization term was applied to the objective function and its parameters were investigated to have the optimal effect on the deformation vector field. Four dimensional computed tomography (4DCT) images from four thoracic and four abdominal cancer patients were used to evaluate the algorithm.
Results: A quarter, half and full of the original voxel resolutions were adopted in the three levels respectively from the coarsest to the finest. An accuracy of 2 mm was achievable for thoracic cancer patients while slightly larger residual error at the average of 3 mm was observed in abdominal cancer patients. Under the same condition, larger errors were found from the calculation with a single resolution model. The smoothness of vector field was significantly improved with the regularization term. The parameters, however, have to be determined empirically. This feature is particularly useful when conducting dose warping.
Conclusion: Significant improvement in DIR accuracy and robustness with the proposed algorithm is achieved. The novel deformable registration model, with more systematic study, will find its widespread application in radiation therapy.