Program Information
Iterative Inversion of Deformation Vector Fields with Feedback Control
A Dubey1*, A Iliopoulos1 , X Sun1 , F Yin2 , L Ren2 , (1) Duke University, Durham, North Carolina, (2) Duke University Medical Center, Durham, NC,
Presentations
MO-F-205-6 (Monday, July 31, 2017) 4:30 PM - 6:00 PM Room: 205
Purpose: To improve both accuracy with respect to inverse consistency and efficiency for iterative inversion of deformation vector fields (DVF), by incorporating data-specific feedback control.
Methods: The presented method introduces and utilizes active feedback control for iterative DVF inversion. At each iteration step, we measure the inconsistency, namely the inverse residual, between the input DVF and the iterative inverse estimate. The residual is modulated by a control mechanism before being fed back into the next iterate. The method includes two previous fixed-point iteration methods as special instances and improves upon them in convergence region and speed by using data-specific parameter for feedback control. The data-specific parameter is determined based on analysis of error propagation, to suppress the inversion error more aggressively, with consideration of the characterization of given dataset. Varying parameter values throughout the iteration were also investigated for further improving convergence. The method was evaluated using both an analytical DVF pair and numerical DVFs obtained from CT images of 7 lung cancer patients.
Results: The data-specific feedback control is analytically and practically shown to attain a larger convergence region at faster pace in iterative DVF inversion, compared to the precursor methods. With the analytical deformation, the iteration converges over the entire domain, and is sped up substantially relative to the precursor methods, which suffer from slow convergence, or even divergence, when displacement is large. With the DVFs from patient CT images, the presented method outperforms the precursor methods in inverse consistency (reducing residual displacement by up to an order of magnitude within 30 iterations), convergence region, and computational efficiency.
Conclusion: Adaptive, data-specific feedback control supports accurate, robust, and efficient iterative DVF inversion, which is valuable to image registration, reconstruction, and adaptive radiotherapy. It also provides a new way of understanding and designing efficient iterative methods for DVF inversion.
Contact Email: