Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study evaluates the impact of different deformation priors in continuous image registration methods, specifically comparing B-Spline and implicit neural representation (INR) approaches. By analyzing performance across inter-subject brain MR and intra-subject lung CT registration tasks, the authors find that the choice of prior significantly influences registration accuracy, with MR-D-BSCP outperforming other methods in challenging deformation scenarios. The results underscore the importance of aligning deformation priors with the characteristics of the target motion, establishing a new principle for medical image registration design.
Matching the right deformation prior to the motion pattern can dramatically enhance medical image registration accuracy, as shown by MR-D-BSCP's superior performance in complex scenarios.
Deformable image registration models implicitly encode deformation priors through their parametrization and optimization. In this work, we conduct a validation study on continuous registration methods to examine how these implicit priors affect performance across different registration tasks. Classic B-Spline transformations impose locality, smoothness, and scale through their control-point structure, whereas recent INR-based methods impose different priors through neural parameterization and optimization. We compare INR-Dense (IDIR), which directly models a dense displacement field using a SIREN-based INR; INR-BSCP (SINR), which predicts B-Spline control points with an INR; D-BSCP, which directly optimizes single-scale B-Spline control points; and MR-D-BSCP, which adds a multiresolution coarse-to-fine scheme. Experiments on inter-subject brain MR registration (OASIS) and intra-subject exhale-to-inhale lung CT registration (DIR-LAB 4DCT) reveal different behavior across deformation regimes. On OASIS, where deformations are moderate but locally complex, D-BSCP matches or slightly outperforms INR-BSCP, suggesting that the B-Spline parameterization accounts for much of INR-BSCP's effectiveness. On DIR-LAB 4DCT, where respiratory motion is larger and more coherent, single-scale B-Spline methods (D-BSCP and INR-BSCP) are less suitable, while INR-Dense and MR-D-BSCP are more effective. Across both tasks, MR-D-BSCP achieves the best performance among the tested continuous parameterizations. These findings highlight that registration accuracy depends strongly on matching the induced deformation prior to the target motion pattern, and support prior-deformation matching as a practical design principle for medical image registration. Our code will be available at https://github.com/HengjieLiu/RightPriorDIR.