Tree-based adaptive finite element methods for deformable image registration
Nicolás A. Barnafi, Alberto F. Martín, Ricardo Ruiz-Baier · 2025
Read Tree-based adaptive finite element methods for deformable image registration by Nicolás A. Barnafi, Alberto F. Martín, Ricardo Ruiz-Baier (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-15876.txt.
Introduction
Nicolás Barnafi, Alberto Martín, and Ricardo Ruiz-Baier propose an adaptive finite element method for deformable image registration. Registration warps one image onto another, a core task in medical imaging. Fixed meshes waste effort where images are smooth and under-resolve where they are not.
The authors pair the method with a residual-based error estimator whose efficiency and reliability they establish theoretically. The estimator guides adaptive mesh refinement and coarsening, concentrating computation where the registration is hardest. The nonlinear equations behind the method are solved with a pseudo time-stepping fixed-point scheme, accelerated by Anderson acceleration. An efficient data structure based on forests of octrees with space-filling curves supports the implementation. Numerical results on application-oriented problems illustrate the performance gains. The work gives imaging scientists a registration solver that spends its budget where the anatomy demands it.
Catalog record
| CREATOR | Nicolás A. Barnafi, Alberto F. Martín, Ricardo Ruiz-Baier |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-15876
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Tree-based adaptive finite element methods for deformable image registration Thanks: Funding: This work has been partially supported by the Monash Mathematics Research Fund S05802-3951284; by the Australian Research Council through the Future Fellowship grant FT220100496 and Discovery Project grant DP22010316; and by the National Research and Development Agency (ANID) of the Ministry of Science, Technology, Knowledge and Innovation of Chile through the grant FONDECYT de Postdoctorado No. 3230326 and the Center for Mathematical Modeling (CMM), Proyecto Basal FB210005. This work was also supported by computational resources provided by the Australian Government through NCI under the ANU Merit Allocation Scheme (ANUMAS). Nicolás A. Barnafi, Alberto F. Martín, Ricardo Ruiz-Baier Abstract In this work we propose an adaptive fem ( fem ) formulation for the dir ( dir ) together with a residual-based a posteriori error estimator, whose efficiency and reliability are theoretically established. This estimator is used to guide amr ( amr ). The nonlinear Euler–Lagrange equations associated with the minimisation of the relevant functional are solved with a pseudo time-stepping fixed-point scheme which is further accelerated using aa ( aa ). The efficient implementation of these solvers relies on an efficient adaptive mesh data structure based on forests-of-octrees endowed with space-filling-curves. Several numerical results illustrate the performance of the proposed methods applied to adaptive dir in application-oriented problems. Keywords : Deformable image registration, a-posteriori error estimates, Anderson acceleration, adaptive mesh refinement and coarsening, forest-of-octrees. Mathematics subject classification (2020) : 65N50, 65N15, 74B20, 74N25. 1 Introduction Image registration problems consist in finding an appropriate mapping of a given template image so that it resembles another reference image. These images can originate from several different applications, and may represent physical parameters, locations, or even time. See [ 35 ] for a review in medical imaging. There are many ways of defining such a problem, including variational formulations [ 10 , 22 ] , level-set methods, grid deformation methods [ 28 ] , learning processes [ 15 ] and Bayesian techniques [ 16 ] , among others. All these approaches differ in how they measure image alignment, with such measure being referred to as similarity . Image registration is named deformable whenever the mapping between images is allowed to vary arbitrarily between pixels, which results in a highly nonlinear problem. Image registration remains a challenging problem in computational imaging due to three inherent difficulties: its ill-posed nature, the selection of appropriate similarity metrics, and the high computational demands of numerical optimisation. First, the problem’s ill-posedness necessitates regularisation strategies, such as elastic potential energy constraints [ 35 ] , to ensure physically plausible solutions. Second, defining effective similarity measures–whether through energy-based criteria or intensity difference metrics like the L^{2} norm–requires careful consideration of image modality and noise characteristics. Third, solving the Euler–Lagrange equations derived from the energy minimisation framework imposes significant computational costs [ 20 ] , particularly when capturing localised deformations. These deformations often exhibit high spatial gradients to resolve fine anatomical or structural details, further amplifying discretisation challenges. Adaptive numerical methods naturally present a promising approach to address these issues. By dynamically refining computational grids or basis functions in regions of sharp deformation, adaptive techniques hold the promise to achieve a remarkable trade-off among accuracy and computational efficiency. Recent advances in amr demonstrate the potential of such strategies to overcome traditional limitations in registration tasks, for which we highlight [ 4 , 20 , 21 , 33 , 42 ] . With this in mind, in this work we leverage hierarchically-adapted non-conforming forest-of-octrees meshes endowed with Morton (a.k.a., Z-shaped) space-filling-curves for storage and data partitioning; see, e.g., [ 1 , 9 ] . These n -cube meshes (made of quadrilaterals or cubes in 2D and 3D, resp.) can be very efficiently handled (i.e. refined, coarsened, re-partitioned, etc.) using high-performance and low-memory footprint algorithms [ 9 ] . However, they are non-conforming at the interfaces of cells with different refinement levels as they yield hanging vertices, edges and faces. In order to enforce the conformity (continuity) of fe ( fe ) spaces built out of this kind of meshes, it is standard practice to equip the space with additional linear multi-point constraints. The structure and set up of such constraints is well-established knowledge, and thus not covered here; see [ 1 ] and references therein for further details. In this work, we consider the dir with an L^{2} similarity measure and a regularisation given by the (linear) elastic energy of the deformation with Robin and pure Neumann boundary conditions. (We note that nonlinear elasticity has also been considered as an alternative regularization term in other works, such as, e.g. [ 18 ] .) This formulation was analysed in [ 3 ] , and an efficient and reliable a-posterior error estimator was developed for it in [ 4 ] for the pure-Neumann case, where uniqueness was obtained by imposing orthogonality against the kernel of the regularisation operator. We extend that analysis to the case of Robin boundary conditions. This efficient and reliable a-posterior error estimator is used at a given solution of the dir on a given mesh in order to automatically adapt it, and the process is repeated across several amr iterations to successively improve the accuracy of the dir solution. As mentioned above, solving dir is a challenging task. It is essentially a nonlinear inverse problem. A well-established strategy to solve it is by means of a pseudo-time formulation [ 31 ] , which is an IMEX approximation of a proximal point algorithm [ 24 ] . In this work, we leverage aa [ 41 ] to improve its robustness and efficiency in terms of the iteration count. Anderson acceleration is a method that considers a fixed-point iteration and yields another one with improved convergence, both in terms of the radius of convergence and the convergence rate [ 36 ] . This technique has already been successfully validated for proximal-point algorithms in the context of inverse problems [ 29 ] , which is the method that most resembles the scheme we use. The main contributions of this work are: (i) the extension of the efficiency and reliability proof of the existing a posteriori error indicator for dir to Robin boundary conditions; (ii) the use of aa to accelerate the pseudo-time solution strategy commonly used in dir ; (iii) the use of the computed a posteriori error estimator to guide efficient amr using octree-based meshes; and (iv) the combination of aa and amr into a novel aa - amr algorithm for solving dir . These methodologies result in significant computational savings and increased accuracy, as shown in our numerical tests. Our software is implemented in the Julia programming language [ 6 ] using the fe tools provided by the Gridap ecosystem of Julia packages [ 37 , 2 , 30 ] . Structure. The remainder of the work has been organised as follows. In Section 2 , we formulate the dir problem. In Section 3 , we propose fem discretisation for dir and its related robust (in the sense of reliability and efficiency) a-posteriori error estimator driving the amr process. In Section 4 , we show how we orchestrate the solution of the nonlinear problem with both aa and amr . In Section