Abstract
The Radon transform is a fundamental tool for analyzing data in tomographic imaging, optimal transport, crystallography, and geometric analysis. Numerical computations require an accurate discretization. To deal with voxelized images and objects, we derive a closed-form, piecewise polynomial expression for the Radon transform of an axis-aligned cube in arbitrary dimension $d$. Building on this formula, we propose a discrete Radon transform in $\mathbb{R}^d$ that is both analytically exact for voxelized data and computationally efficient. For improved numerical stability, we introduce a regularized variant replacing the Radon transform of a cube, i.e. the $(d-1)$-dimensional area of the intersection between that cube and a hyperplane, by the $d$-dimensional volume of the intersection between the cube and a thin slab around the hyperplane. Numerical experiments demonstrate the effectiveness of the proposed approach in several applications including 3D shape matching, classification, and sliced Wasserstein barycenters. The computational efficiency in higher dimensions is verified by a comparison with Monte Carlo integration.
Introduction
The Radon transform is a mathematical workhorse used to reconstruct 3D objects from 2D projections or to analyze geometric shapes. However, while the transform is defined for continuous functions, empirical data usually consists of discrete "voxels" (3D pixels). Standard discretization methods often suffer from inaccuracy (ignoring voxel volume), artifacts (Fourier-based ringing), or high computational costs (Monte Carlo sampling). This paper addresses the fundamental geometric challenge: exactly calculating the area of a plane as it slices through a single 3D cube at any arbitrary angle and offset.
Methodology
The core contribution is a "voxel-aware" discrete Radon transform (DRT). By deriving an exact, closed-form formula for the intersection area between a hyperplane and an axis-aligned hypercube, the authors replace slow simulations and inaccurate approximations.
Analytical Intersection Formula
The paper introduces Theorem 1, providing a closed-form, piecewise polynomial expression for the Radon transform of a hypercube in any dimension $d$. This is achieved by recognizing that the Radon transform of a cube can be expressed as a convolution of indicator functions, relating directly to polynomial B-splines.
The discrete Radon transform of a voxelized image becomes a weighted sum where each voxel's contribution is its exact intersection area with the integration plane. Unlike previous geometric methods requiring complex triangulation, this formula uses a summation over the $2^d$ corners of the cube, enhancing efficiency in low to moderate dimensions.
Regularization via Slab Intersection
To improve numerical stability, the authors derive a "slab" formula. This replaces the infinitely thin plane with an $\varepsilon$-neighborhood (a thin slab), calculating the $d$-dimensional volume of the intersection. This acts as a natural regularizer, reducing sensitivity to the discretization grid's exact placement.
Experiments and Results
The method was evaluated using the ModelNet10 dataset (voxelized 3D CAD models) and synthetic empirical measures for optimal transport tasks.
Computational Efficiency
In 4D experiments, evaluating the exact formula for 128 angles and 128 radii took less than 1 second. In contrast, Monte Carlo sampling required approximately 5 seconds to reach even a 1% error margin.
Shape Classification and Invariance
Using the max-normalized Radon cumulative distribution transform (mNR-CDT), the approach achieved 93% accuracy on a 3-class subset of ModelNet10 and 47% on the full 10-class set, outperforming previous Radon-based extractors. Furthermore, the method demonstrated high robustness to affine transformations, maintaining 99% accuracy under random scaling, shearing, and rotation.
Optimal Transport
The paper successfully computed Sliced Wasserstein barycenters for 3D measures, effectively "blending" complex 3D shapes. This demonstrates the utility of the exact DRT in high-dimensional data analysis.
Conclusion
This work provides a faster and more precise tool for 3D shape analysis by leveraging an analytical solution for cube-plane intersections. While the $2^d$ complexity may limit extremely high-dimensional applications, the efficiency in 3D and 4D is significant. Future work includes GPU acceleration and applying this exact discretization to inverse reconstruction problems in medical imaging.
Summary updated