Source-linked AI summary
Tensor-based formulation and nuclear norm regularization for multi-energy computed tomography
Oguz Semerci, Ning Hao, Misha E. Kilmer, Eric L. Miller
TL;DR
Multi-energy CT seeks energy-dependent attenuation reconstructions enabled by photon-counting detectors, but requires regularization that uses the multi-spectral structure. The paper models the unknown as a low-rank 3-way tensor and develops tensor nuclear-norm regularizers solved iteratively. Simulations show that these regularizers reduce noise, with combinations of tensor nuclear norms and TV particularly improving low-energy images.
Problem
Photon-counting detectors enable energy-selective CT, motivating reconstruction methods that recover energy-dependent attenuation while exploiting multi-spectral structure.
Method
The paper models the multi-spectral unknown as a low-rank 3-way tensor and develops generalized and t-SVD-based tensor nuclear-norm regularizers for iterative reconstruction.
Results
The tensor nuclear norm works as a stand-alone regularizer, while combining it with TV enhances regularization, especially for noisy low-energy images.
Takeaways & Limitations
Tensor nuclear-norm regularization provides a framework for exploiting multilinear structure in multi-energy CT and other multilinear inverse problems.
Takeaways & Limitations
Practical use remains limited by the perfect-energy-resolution assumption, computational cost, and the need for testing with experimental data and higher-resolution reconstructions.
Abstract
from arXiv · showhide
The development of energy selective, photon counting X-ray detectors allows for a wide range of new possibilities in the area of computed tomographic image formation. Under the assumption of perfect energy resolution, here we propose a tensor-based iterative algorithm that simultaneously reconstructs the X-ray attenuation distribution for each energy. We use a multi-linear image model rather than a more standard "stacked vector" representation in order to develop novel tensor-based regularizers. Specifically, we model the multi-spectral unknown as a 3-way tensor where the first two dimensions are space and the third dimension is energy. This approach allows for the design of tensor nuclear norm regularizers, which like its two dimensional counterpart, is a convex function of the multi-spectral unknown. The solution to the resulting convex optimization problem is obtained using an alternating direction method of multipliers (ADMM) approach. Simulation results shows that the generalized tensor nuclear norm can be used as a stand alone regularization technique for the energy selective (spectral) computed tomography (CT) problem and when combined with total variation regularization it enhances the regularization capabilities especially at low energy images where the effects of noise are most prominent.
1 Introduction
Multi-energy CT uses photon-counting detectors to recover energy-dependent attenuation, motivating a tensor-based reconstruction that exploits multilinear structure through convex nuclear-norm regularization. The proposed tensor regularizers address limitations of conventional tensor decompositions and improve low-energy reconstruction when combined with total variation.
- Motivation: Photon-counting detectors distinguish photon energies, enabling spectral attenuation recovery and potentially improving material characterization in medical and security imaging.Unlike energy-integrating detectors, photon-counting detection also avoids the main source of detector read-out noise.
- Problem formulation: The authors model the multi-spectral unknown as a low-rank 3-way tensor rather than a matrix of vectorized energy images.The tensor dimensions represent two spatial dimensions and energy.
- Motivation: Standard CP and Tucker decompositions require difficult non-convex optimization and prior knowledge of the number of components.These convergence and model-selection issues motivate nuclear-norm-based alternatives.
- Tensor regularization: The generalized tensor nuclear norm sums nuclear norms across tensor unfoldings, extending matrix nuclear-norm regularization to multilinear inverse problems.The paper identifies this formulation as a first contribution and assumes low rank in each unfolding.
- Tensor regularization: A second regularizer, TNN-2, uses tensor singular value decomposition to account for correlations across all tensor dimensions.The t-SVD-based norm can also be expressed in matrix nuclear-norm form.
- Results: Combining tensor nuclear norms with total variation improves reconstruction quality especially at low energies, where reduced photon counts produce more prominent noise.The tensor prior couples information across energies while TV regularizes each energy image spatially.
2 Preliminaries on Tensors
This section introduces tensors, their unfoldings, and the operations needed for the t-SVD-based tensor nuclear norm regularizer. It defines both conventional mode-wise rearrangements and the t-product framework for 3-way tensors.
- Tensor notation: A K-way tensor is a multilinear structure in R^{N1×N2×⋯×NK}, generalizing two-dimensional matrices to multiple dimensions.The paper uses χ for tensors and X^(k) for frontal faces of a 3-way tensor.
- Tensor notation: Mode-l unfolding rearranges tensor elements into a matrix with mode-l fibers as columns, using a corresponding permutation relationship for vectorized forms.For 3-way tensors, unfoldings can be visualized through frontal, horizontal, and lateral slices.
- Tensor operations: The fold and unfold operators convert between tensor form and an arrangement in which frontal slices are aligned, enabling subsequent block-circulant constructions.The paper states that fold(unfold(χ)) returns the tensor χ.
- Tensor operations: The n-mode product multiplies a tensor by a matrix along one selected dimension, producing a tensor whose corresponding mode has the matrix’s row dimension.This operation is defined for a K-way tensor and is distinct from multiplication of two 3-way tensors.
- Tensor operations: The t-product multiplies compatible 3-way tensors and produces a tensor in R^{N1×ℓ×N3}; it forms the basis for the t-SVD and TNN-2 regularizer.The t-product is defined through fold, block-circulant construction, and unfold operations.
- Tensor operations: The t-SVD framework uses tensor transpose, orthogonality under the t-product, block diagonalization, and Fourier-domain frontal faces.The Fourier-domain representation applies the FFT to mode-3 fibers and supports the t-SVD construction.
3 The Measurement Model and The Multi-spectral Unknown as a Tensor
The measurement model describes energy-resolved CT under parallel-beam acquisition and perfect energy resolution, then organizes the energy-specific images into one 3-way tensor. This unified representation supports reconstructing all energy channels jointly rather than independently.
- Measurement model: Polychromatic CT models measurements using the energy-dependent attenuation coefficient, source spectrum, and X-ray path; perfect energy resolution simplifies each energy-bin model to a monochromatic projection.The work uses parallel-beam measurement geometry and assumes infinitesimal detector-bin width.
- Measurement model: The model contains N_3 energy bins, with k indexing detector bins and j indexing source-detector pairs.The paper notes that electronic noise can be neglected for photon-counting detectors, unlike conventional energy-integrating CT.
- Measurement model: The discrete model represents each energy-dependent attenuation distribution as an image vector x_k with N_p = N_1N_2 pixels and uses a system matrix A for ray-path lengths through pixels.The acquisition is discretized into N_m source-detector pairs, with measurements incorporating Poisson statistics.
- Tensor representation: The multi-spectral unknown is represented as χ ∈ R^{N_1×N_2×N_3}, whose first two dimensions are spatial and whose third dimension is energy.The energy-specific vectors x_k correspond to the lexicographically ordered frontal slices of χ.
- Tensor representation: The reconstruction goal is to estimate χ from measurements y_{k,j} while treating all energy-specific images in a unified manner rather than reconstructing them independently.The multi-linear structure can also be extended to higher-dimensional inverse problems, although this paper focuses on multi-energy CT.
4 Low-Rank modeling and Regularization
The paper extends low-rank modeling from matrices to multi-energy CT tensors using convex tensor nuclear norms based on unfoldings and t-SVD. These regularizers exploit correlations across spatial and energy dimensions without requiring tensor decomposition ranks in advance.
- Matrix low-rank modeling: Low-rank modeling supplies regularization and prior information for multidimensional inverse problems, but direct rank minimization is non-convex and NP-hard.The matrix nuclear norm is used as a convex relaxation of rank minimization.
- TNN-1: The generalized tensor nuclear norm sums the nuclear norms of tensor unfoldings across modes, extending matrix nuclear-norm minimization to tensors.Unfolding rearranges tensor columns along a selected mode into a matrix.
- TNN-1: TNN-1 assumes the multi-spectral unknown is low-rank in each unfolding, thereby exploiting correlations across energy and spatial dimensions.The energy unfolding is expected to be low-rank because neighboring energies are highly correlated, while spatial unfoldings can capture structural redundancies.
- Empirical low-rank structure: The singular values of both the tensor unfoldings and bcirc(χ) decay rapidly in phantom examples, indicating the usefulness of the corresponding nuclear-norm regularizers.Figure 5 displays unfolding singular values, while Figure 6 displays block-circulant singular values.
- TNN-2: TNN-2 uses the t-SVD and defines its tensor nuclear norm through the nuclear norm of the block circulant matrix bcirc(χ).The t-SVD supports truncated optimal low-rank representation in Frobenius norm, and TNN-2 uses one regularization parameter.
5 Inverse Problem Formulation
The inverse problem is formulated as a convex penalized weighted least-squares objective derived from a quadratic approximation to the Poisson log-likelihood. Tensor nuclear-norm regularization is combined with either slice-wise TV or 3D-TV regularization.
- Data fidelity: The measurement model yields a penalized weighted least-squares formulation that quadratically approximates the Poisson log-likelihood for each energy bin.The data fidelity term is defined separately for the kth energy bin before the regularizer is added.
- Regularized objective: The resulting convex objective combines a tensor nuclear-norm regularizer with total variation regularization.The tensor term may be R∗(χ) or R⊛(χ).
- TV choices: The slice-wise TV operator applies weighted isotropic 2D TV to the frontal energy slices, with energy-specific regularization parameters αk.The alternative 3D-TV operator extends the gradient across all three tensor dimensions.
- Regularization behavior: TV favors sparse gradients and piecewise-constant reconstructions but can produce stair-casing that is problematic for texture recovery.The paper reports that combining tensor nuclear norm with TV reduces the TV required for noise cancellation and helps recover texture.
6 Solution Algorithm via alternating direction method of multipliers (ADMM)
ADMM solves the convex reconstruction problem by splitting it into tractable subproblems and alternating updates of the image, auxiliary variables, and dual variables. The resulting updates include singular-value shrinkage and TV-regularized quadratic solves.
- ADMM framework: ADMM decomposes the difficult objective into tractable pieces through variable splitting, although it increases the number of minimization variables.The paper describes this as a combination of dual decomposition and augmented Lagrangian methods.
- TNN-1 updates: For the TNN-1 formulation, the optimization is reformulated with an augmented Lagrangian and solved by alternating updates of χ, auxiliary variables Zk, and dual variables.The χ and Zk updates are performed alternately before updating the dual variables.
- Energy-wise updates: Permutation notation decouples the χ update by energy, allowing each xk to be treated separately as a TV-regularized quadratic problem.The paper solves these subproblems using FISTA.
- Nuclear-norm proximal step: The auxiliary nuclear-norm subproblem has an analytical singular-value-shrinkage solution, shrinking each singular value by the threshold ρ.The shrinkage operator uses (σi−ρ)+, where t+=max(t,0).
- TNN-2 updates: The TNN-2 algorithm follows the same solution structure as TNN-1 after replacing TNN-1 with TNN-2.The block-circulant formulation is used to express the corresponding updates.
7 Reconstruction Examples
The simulations compare FBP, TV, TNN, and combined regularizers across multi-energy CT phantoms, evaluating relative ℓ2 error and computation time. Tensor nuclear norms reduce noise, preserve detail, and improve low-energy reconstructions, with combined methods offering stronger quality at greater computational cost.
- Methods and evaluation: Relative ℓ2 error measures reconstruction accuracy, while reported tables also track computation time.Figure 10 plots relative ℓ2 error against iteration number and compares it with the FBP error level.
- Results: FBP fails to provide reasonable low-energy reconstructions, whereas the proposed methods outperform FBP within at most 3 iterations for 85 keV Phantom-1.The low-energy failure is attributed to limited numbers of views and noise.
- Results: TNN-1 and TNN-2 alone provide considerable noise reduction while preserving much detail and allow rapid processing relative to the other methods.The regularization parameters were manually tuned and kept the same for both phantoms, with stronger regularization assigned to low-energy images.
- Results: Combining TNN-1 or TNN-2 with TV improves detail preservation and low-energy reconstruction quality, but increases computational burden.These combinations produce superior results to 3D-TV, which incorporates smoothness along the energy dimension.
- Results: TNN-1 and TNN-2 perform similarly as sole regularizers, while TNN-2 outperforms TNN-1 for 85 keV images when combined with TV.Increasing the number of energy bins from 12 to 25 still leaves spatial unfoldings essential to the regularization.
8 Conclusions
The paper concludes that generalized tensor nuclear norms provide useful regularizers for iterative multi-energy CT. It also identifies extensions and practical work needed before broader deployment, including automatic parameter selection, faster implementations, and testing on experimental data.
- Conclusions: The study provides an algorithmic framework for iterative multi-energy CT using generalized tensor nuclear norm regularizers.It also proposes an alternative tensor nuclear norm based on t-SVD.
- Conclusions: The tensor nuclear norm framework can extend to inverse problems with multilinear unknowns and to higher-dimensional settings such as 5D data.The proposed extensions include an additional spatial dimension and time dependency.
- Future work: Future work includes low-rank-sparse models, Tucker and CANDECOMP/PARAFAC decompositions, and estimating redundancy across tensor dimensions.These directions are intended to support more sophisticated tensor nuclear norm regularizers.
- Motivation: Tensor regularization is motivated by structurally similar energy images and may reduce the stair-casing and cartooning effects associated with TV.The paper emphasizes this motivation particularly for medical imaging applications.
- Limitations and future work: Automatic regularization-parameter selection, reduced computation time, parallel or optimized implementations, and tests with real experimental data remain necessary for practical use.The paper also calls for higher-resolution reconstructions and real-life scenario testing.