Test-Time Adaptation for Deformable Image Registration
Y Sang*, M McNitt-Gray, Y Yang, M Cao, D Low, D Ruan, University of California, Los Angeles, CA
Presentations
TU-I345-IePD-F5-6 (Tuesday, 7/12/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 5
Purpose: In classic optimization-based registration method, the deformation vector field (DVF) is obtained for each image pair by solving the corresponding optimization problem which typically involves time consuming iterative processes. Deep-learning-based registration offers a much faster alternative and can benefit from data-driven regularizing behaviors. However, training and testing samples can differ in either image or motion characteristics or both, raising concerns about lack of generalization reliability in specific individual test case from direct inference. In this study, we propose a domain adaptation method to address the potential domain shift, and to improve the accuracy and robustness of registration.
Methods: Our method utilizes a descriptor network to impose general feasibility prior on DVF. The trained registration network is further adapted for each image pair at test time to optimize the individualized performance. The adaptation method was tested under various generalization domain shifts in cross-protocol, cross-platform, and cross-modality scenarios, with test evaluation performed on nine lung CBCTs, 15 cardiac MRIs, and one lung MRI, respectively. Our method was compared against a manually tuned classic B-spline method in SimpleElastix and the network without adaptation.
Results: Our method achieved (2.11 +/- 1.61), (2.26 +/- 1.41), and (2.00 +/- 1.63) mm landmark-based registration errors on lung CBCT, cardiac MRI, and lung MRI, respectively. In a motion-compensated CBCT enhancement test, it achieved (102.1 +/- 7.96) HU root-mean-square error and (0.994 +/- 0.002) structural similarity index to the ground-truth CT. Our method improved the registration accuracy with respect to individual test input, with statistical significance.
Conclusion: We have demonstrated the efficacy to adjust trained registration network to unseen data acquired on a different protocol or scanner, with a different imaging platform, or even modality. The proposed paradigm enables applying deep learning registration method to less common imaging types or diseases when the data is limited.
Keywords
Taxonomy
IM/TH- Image Registration: General (Most aspects)
Contact Email