Comparative analysis of numerical methods for X-ray 3D reconstruction from sparse views
Abstract
Reconstruction of three-dimensional vascular structures from a limited number of X-ray projections is a severely underdetermined and ill-conditioned inverse problem. This work presents a unified matrix-free computational framework for evaluating iterative numerical methods for sparse-view cone-beam tomography. Forward and backprojection are implemented from ray-voxel intersection lengths computed with the Amanatides-Woo traversal algorithm, avoiding explicit construction of the large sparse system matrix. We systematically compare ART, SART, SIRT, gradient descent (GD), and Krylov subspace methods, including Conjugate Gradient, LSQR, and GMRES. The common matrix-free operators and geometry are used for all methods, allowing consistent comparisons under matched conditions. Particular emphasis is placed on gradient scaling using scalar step sizes, Jacobi preconditioning, and diagonal Hessian majorization, together with L1, Tikhonov, and total variation regularization. Numerical experiments use five three-dimensional coronary vessel models reconstructed from ten simulated X-ray projections under noise and scattering conditions, and are extended to real X-ray projections of a static thin-vessel phantom. Performance is evaluated in terms of convergence, computational cost, volumetric overlap, contrast-to-noise ratio, vessel diameter and centerline errors, and projection consistency at an unseen viewing angle. The results show that total-variation-regularized gradient descent with diagonal majorization provides the best overall reconstruction performance within the investigated computational budgets. By contrast, the tested Krylov implementations achieve limited geometric recovery despite higher computational cost. These findings emphasize the importance of appropriate gradient scaling and regularization for stable and computationally efficient sparse-view tomographic reconstruction.
Related
- Rapid reconstruction-free classification from Lissajous sparse sampling using deep learning
- Reconstruction of Binary Linear Systems and Profile Geometry of Sparse Krylov Strata
- Numerical reconstruction of polygonal obstacles by diffracted waves
- Numerical methods for nonlinear stochastic differential equations driven by subdiffusion
- Well-posedness and numerical reconstruction of a source term for linear parabolic problems with an integral constraint
BibTeX
@misc{bardi2026comparative,
title = {{Comparative analysis of numerical methods for X-ray 3D reconstruction from sparse views}},
author = {Melitine N. Bardi and Konstantinos K. Delibasis},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.05645},
eprint = {2611.05645},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.05645},
note = {CC BY-NC-SA 4.0},
}
RIS
TY - UNPB
TI - Comparative analysis of numerical methods for X-ray 3D reconstruction from sparse views
AU - Melitine N. Bardi
AU - Konstantinos K. Delibasis
PY - 2026
DA - 2026/11/03
PB - arXiv
DO - 10.48550/arxiv.2611.05645
UR - https://arxiv.org/abs/2611.05645
AB - Reconstruction of three-dimensional vascular structures from a limited number of X-ray projections is a severely underdetermined and ill-conditioned inverse problem. This work presents a unified matrix-free computational framework for evaluating iterative numerical methods for sparse-view cone-beam tomography. Forward and backprojection are implemented from ray-voxel intersection lengths computed with the Amanatides-Woo traversal algorithm, avoiding explicit construction of the large sparse system matrix. We systematically compare ART, SART, SIRT, gradient descent (GD), and Krylov subspace methods, including Conjugate Gradient, LSQR, and GMRES. The common matrix-free operators and geometry are used for all methods, allowing consistent comparisons under matched conditions. Particular emphasis is placed on gradient scaling using scalar step sizes, Jacobi preconditioning, and diagonal Hessian majorization, together with L1, Tikhonov, and total variation regularization. Numerical experiments use five three-dimensional coronary vessel models reconstructed from ten simulated X-ray projections under noise and scattering conditions, and are extended to real X-ray projections of a static thin-vessel phantom. Performance is evaluated in terms of convergence, computational cost, volumetric overlap, contrast-to-noise ratio, vessel diameter and centerline errors, and projection consistency at an unseen viewing angle. The results show that total-variation-regularized gradient descent with diagonal majorization provides the best overall reconstruction performance within the investigated computational budgets. By contrast, the tested Krylov implementations achieve limited geometric recovery despite higher computational cost. These findings emphasize the importance of appropriate gradient scaling and regularization for stable and computationally efficient sparse-view tomographic reconstruction.
ER -