Low-dose spectral CT reconstruction using L0 image gradient and tensor dictionary

1. Introduction

• Dual-energy CT provides only two attenuation measurements, limiting the material decomposition maps that can be discriminated.The introduction also notes spectral overlap from energy-integrating detectors.Source: §1 · p. 1
• Photon-counting detector cross-talk and energy-window overlap can lower measurement SNR and compromise material decomposition.The introduction identifies accurate, clean volumetric material images as a primary spectral CT challenge.Source: §1 · p. 2
• The gradient ℓ0-norm penalizes the number of nonzero gradients rather than their magnitudes, supporting edge-direction preservation and fine-structure recovery.The authors state that this spatial sparsity can reduce artifacts and improve denoising.Source: §1 · p. 3
• ℓ0TDL combines tensor dictionary learning with image-gradient ℓ0-norm minimization to address TDL’s weakness in preserving edge information.The gradient constraint also emphasizes spatial sparsity, which the authors say can reduce artifacts and improve denoising.Source: §1 · p. 3
• The work also develops an ADMM solver and evaluates ℓ0TDL through numerical simulations and preclinical mouse studies.Its stated advantages include recovering fine structures and edge information while suppressing noise.Source: §1 · pp. 3–4

2.C. Image gradient ℓ0-norm minimization

• The method considers third-order tensors for its tensor-based dictionary-learning formulation.The dictionary-learning objective constrains the sparse coefficient representation by a sparsity level.Source: §2 · p. 4

2.A. Tensor dictionary learning

• Tensor dictionary learning represents tensor data with a dictionary and sparse coefficients, constrained by a sparsity level.The method alternates coefficient and dictionary updates; MOMP updates the sparse coefficients while the dictionary is fixed.Source: §2 · p. 4

2.B. Tensor dictionary learning for spectral CT

• The spectral CT model treats reconstructed images and projection data as third-order tensors across energy channels and fits the projections using a system matrix.Its objective also includes tensor-patch dictionary representation and regularization terms.Source: §2 · p. 4

2.C. Image gradient ℓ0-norm minimization

• The image-gradient ℓ0-norm counts locations with nonzero gradients, rather than penalizing gradient magnitudes.The method states that this can preserve edge information and fine structures.Source: §2 · p. 5

3.A. ℓ0TDL mathematic model

• The proposed reconstruction framework combines image-gradient ℓ0-norm minimization with tensor dictionary learning to target more accurate edge information and less noise.The model is motivated by TDL’s difficulty recovering high-quality edges from severely noisy, undersampled low-dose spectral CT images.Source: §3 · pp. 5–6

3.B. Solution

• Because the objective includes a non-convex, NP-hard gradient ℓ0 term, the method uses ADMM and split-Bregman updates to solve for its variables iteratively.An auxiliary variable is introduced, and the updates alternate among the reconstruction and auxiliary variables.Source: §3 · p. 6
• The ℓ0TDL workflow trains a global tensor dictionary with K-CPD, then iteratively updates the reconstruction, auxiliary variables, sparse codes, and dictionary.The workflow also applies a positive constraint to the reconstruction and denormalizes the final image.Source: §3 · p. 8

3.C. Selection of parameters

• The method sets K=1024 for the dictionary and identifies sparse level L, precision level ε, and smoothness control factor λ* as optimization parameters.Regularization parameters σ and η balance the two regularization programs.Source: §3 · p. 9

Regularization parameters σ and η

• Larger σ raises SSIM and FSIM, especially in high-energy channels, but can blur images and remove finer structures; RMSE differences across σ values are small.The σ-dependent image-quality assessments appear in Fig. 2, and representative reconstructions appear in Fig. 3.Source: §3 · p. 9, Figure 2 · p. 9, Figure 3 · p. 10
• The parameter η has little image-quality impact compared with σ, and the two regularization parameters are interdependent.With the image-gradient ℓ0-norm term fixed, η can be adjusted empirically to reach an optimization solution.Source: §3 · p. 9, Figure 4 · p. 10

TDL parameters ε and L

• The parameter ε plays an important role in controlling reconstructed images: greater ε can improve image quality to some extent but lose finer structures, while smaller ε may perform poorly in denoising and produce salt-and-pepper noise.Source: §3 · p. 10
• A lower sparsity level L better protects edge information, whereas a higher level may remove finer structures.The assessment results for different L values are presented in Fig. 6.Source: §3 · p. 11, Figure 6 · p. 11

Smooth factor λ*

• A greater λ* improves image quality and protects image edges, but λ* cannot be increased without limitation; otherwise, the reconstruction becomes oversmoothed and edge information is lost.The parameter λ* is designed to control the smoothness of the reconstructed image.Source: §3 · p. 11

4.A.1 Sparse-view image reconstruction

• Compared with FBP, TV, TV+LR and TDL, the algorithm retained image edge information and had good anti-noising performance in modified mouse-thorax phantom reconstructions from 106 and 80 projections.Source: Figure 9 · p. 13, Figure 10 · p. 13, §4 · p. 14
• ℓ0TDL achieved the smallest RMSE in the 1st, 4th, and 8th channels across all sparse-view projection counts; TV+LR ranked next in lower-energy channels and TDL in higher-energy channels.TV minimization had the largest RMSE in all channels.Source: Table 2 · p. 14, §4 · p. 14
• For bone, ℓ0TDL mean values were most accurate, with relative biases below 1.8% in all channels; for iodine, ℓ0TDL relative biases were below 1.6%.Source: §4 · p. 15
• For 80-view material decomposition, ℓ0TDL gave the most accurate bone components and finer soft-tissue structures; its iodine accuracy was similar to TDL.TV, TV+LR, and TDL misclassified the bony region.Source: §4 · p. 16, Figure 13 · p. 17
• ℓ0TDL converged quickly to a lower-RMSE solution than competing algorithms, with its RMSE decreasing rapidly before stabilizing after 40 iterations.The comparison uses average RMSE versus iteration number.Source: §4 · p. 17, Figure 14 · p. 17
• For 80 projections, ℓ0TDL took 10.68 seconds per iteration, longer than TV (4.60), TV+LR (4.70), and TDL (5.54), due to gradient ℓ0-norm minimization.The gradient optimization is implemented channel by channel.Source: §4 · p. 18

Low-dose reconstruction

• For low-dose projection datasets with 80 views, the proposed ℓ0TDL method can provide much finer structures compared with the other reconstruction methods.Source: 4.A.2 Low-dose reconstruction · p. 18
• ℓ0TDL achieves the minimum RMSE and maximum SSIM and FSIM among competing methods on low-dose data; FBP has greater average RMSE than iterative methods.For 4×10^3 photons, TV has the highest RMSE among iterative methods, followed by TDL and TV+LR.Source: 4.A.2 Low-dose reconstruction · p. 19, Table 3 · p. 19
• ℓ0TDL obtains the minimum RMSEs for all basis materials in low-dose reconstructions.RMSEs are assessed against reference images reconstructed from normal-dose, noise-free projections.Source: 4.A.2 Low-dose reconstruction · p. 19, Table 4 · p. 19

Real mouse sparse-view reconstruction

• In the 120-view real mouse reconstruction, ℓ0TDL best protects image edges and reduces x-ray beam-hardening artifacts.The authors attribute the possible artifact reduction to penalizing sparsity in the image-gradient domain with the gradient ℓ0-norm.Source: 4.A.2 Low-dose reconstruction · p. 20, Figure 17 · p. 20
• The proposed algorithm achieves the most accurate material decomposition compared with other competing techniques, while ℓ0TDL’s rendered color image shows clearer edges than those of other algorithms.Source: 4.A.2 Low-dose reconstruction · p. 20
• FBP and TV reconstructions contain severe artifacts; TV smears finer structures, TV+LR cannot distinguish arrow-indicated structures, and TDL’s yellow-arrow region has severe artifacts that make the bone edge difficult to distinguish.Source: 4.A.2 Low-dose reconstruction · p. 21
• With 40 real-data projections, ℓ0TDL preserves bone boundaries, while TV+LR can incorrectly decompose bone material.Source: 4.A.2 Low-dose reconstruction · p. 22, Figure 20 · p. 22

• Numerical simulations and realistic preclinical mouse studies show that ℓ0TDL outperforms TV, TV+LR, and TDL, with improvement especially in low-dose and sparse-view reconstruction.Source: §5 · p. 22
• The method also suppresses ring artifacts around the bone region and performs strongly in sparse-view reconstruction.Source: §5 · p. 22
• At 80 views, ℓ0TDL better recovers fine structures and protects edges than TV+TDL in simulations, with realistic experiments showing consistent conclusions.Source: §5 · p. 23
• The numerous regularization and control parameters were selected empirically, while their theoretical analysis and optimization remain open problems.Source: §5 · p. 24
• The global tensor dictionary is trained from full-projection FBP priors; when unavailable, fallback training may compromise image quality, while adaptive updates cost more computation.Source: §5 · p. 24
• The conclusion credits ℓ0TDL with maintaining fine structures and image edges while reducing beam-hardening artifacts.Source: §5 · p. 24