TV-LoRA for Sparse-View CT

• TV-LoRA is a hybrid CT reconstruction framework that integrates a diffusion generative prior, anisotropic TV, and low-rank (nuclear norm) constraint to accurately restore images from severely undersampled data.
• It utilizes an ADMM optimization loop with PCG–FFT acceleration, balancing data consistency with learned anatomical priors to preserve fine textures and sharp edges.
• Experimental results demonstrate significant improvements in PSNR and SSIM over traditional methods, effectively suppressing artifacts in extremely sparse-view and low-dose CT settings.

TV-LoRA, short for Diffusion–Low-Rank Hybrid Reconstruction for Sparse-View Medical Imaging, is a reconstruction framework for sparse-view and low-dose CT (LDCT) designed to recover high-fidelity CT images from severely undersampled projection data, where standard inverse solvers become ill-posed and purely learning-based methods may hallucinate anatomically incorrect content. It targets regimes with 8, 4, or even 2 projection angles, and combines a diffusion generative prior, anisotropic total variation (TV), and a low-rank LoRA / nuclear-norm regularizer within an ADMM optimization loop, with the linear subproblem accelerated by PCG + FFT-style frequency-domain structure exploitation. The stated objective is to preserve fine texture, sharp edges, and suppress streak artifacts while maintaining tractable inference for large-scale CT reconstruction (Deng et al., 5 Oct 2025).

1. Problem setting and conceptual scope

The reconstruction problem is posed under the standard sparse-view CT forward model

$b \approx Ax,$

where $x$ denotes the unknown CT image, $A$ the system matrix or projection operator, and $b$ the measured sinogram. In this setting, severe view undersampling makes the inverse problem ill-posed, especially under low-dose conditions and at extreme sparsity such as 8, 4, or 2 projection angles.

TV-LoRA is presented as a hybrid framework that integrates three complementary ingredients. The first is a diffusion generative prior, intended to inject realistic anatomy and denoise the reconstruction. The second is anisotropic total variation, used to preserve local piecewise-smooth structure and sharp boundaries. The third is a low-rank LoRA / nuclear-norm regularizer, intended to capture global patch redundancy and texture consistency. These components are coupled through ADMM, so that generative and physical constraints act jointly rather than as isolated post-processing stages (Deng et al., 5 Oct 2025).

A central point in the formulation is that the method is neither a purely model-based inverse solver nor a purely learned image generator. The paper explicitly motivates TV-LoRA by the failure modes of both classes: standard inverse solvers struggle with severe undersampling, while purely learning-based methods may hallucinate anatomically incorrect content. This suggests that the framework is best understood as a constrained reconstruction method in which learned priors are subordinated to data consistency and explicit regularization.

2. Variational formulation and constituent priors

The hybrid reconstruction objective is written as

$$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$

Here, $\frac{1}{2}\|Ax-b\|_2^2$ enforces data consistency, $-\log p_\theta(x)$ is the diffusion prior, $\mathrm{TV}(x)$ is anisotropic TV, and $\|P(x)\|_*$ is the nuclear norm of a patch-unfolded image matrix.

The role of $P(x)$ is explicit: the image is divided into patches and reshaped into a block matrix, and the nuclear norm encourages those patch matrices to be low rank. In the paper’s terminology, this low-rank component is called LoRA regularization. A common misconception is to interpret “LoRA” here through the lens of low-rank adaptation for parameter-efficient fine-tuning. In TV-LoRA, however, the term refers mathematically to a nuclear norm penalty on patch-wise unfolded image blocks, not to parameter adaptation in the diffusion model. The paper states this directly by describing the term as LoRA / nuclear norm low-rank regularization (Deng et al., 5 Oct 2025).

The rationale for combining these priors is structurally differentiated. Anisotropic TV targets local regularity and boundary preservation; the low-rank nuclear norm targets repeated structures and inter-patch correlation; the diffusion prior supplies a learned anatomical model capable of restoring realistic CT structure from noisy or incomplete estimates. The paper’s ablation results, discussed below, support the claim that these effects are complementary rather than redundant.

3. Diffusion prior: score-based modeling with NCSN++

The diffusion module is a score-based diffusion model based on NCSN++. The forward process is defined as a variance-exploding SDE in which a clean image $x$0 is progressively perturbed by Gaussian noise: $x$1 with terminal distribution

$x$2

The implementation uses a VESDE schedule with

$x$3

and $x$4 diffusion steps.

The reverse-time dynamics are parameterized by a score network $x$5, leading to

$x$6

In discrete form, the paper gives the Euler–Maruyama update

$x$7

This reverse SDE constitutes the learned generative prior used during reconstruction.

The NCSN++ backbone is specified as a four-level U-Net with channel widths

$x$8

using Conv–GroupNorm–SiLU blocks, skip connections, and a GEGLU global attention module at the bottleneck to better model long-range texture dependencies. Training uses denoising score matching with a multi-scale $x$9 residual loss, Adam with learning rate $A$0, EMA decay 0.999, and cosine annealing. The model contains about 43.7M parameters (Deng et al., 5 Oct 2025).

Within the overall framework, the diffusion module is not presented as a stand-alone reconstructor but as one term in a constrained optimization procedure. That distinction matters: the paper attributes noise suppression, hallucination avoidance, and structural realism to the diffusion prior, but only in concert with data consistency and explicit regularizers.

4. TV and low-rank regularization within ADMM

To preserve edges and suppress local noise and artifacts, TV-LoRA introduces anisotropic TV through auxiliary gradient variables

$A$1

The TV penalty is

$A$2

which in the constrained form becomes

$A$3

This term encourages piecewise smoothness while preserving sharp anatomical boundaries.

The low-rank component is expressed as

$A$4

with auxiliary variable

$A$5

Its motivation is the observation that medical CT images contain repeated structures and strong inter-patch correlations. The paper states that low-rankness helps recover missing textures and global consistency, complementing TV’s local edge-preserving behavior.

The constrained optimization problem is rewritten as

$A$6

subject to

$A$7

It is then solved using ADMM with scaled multipliers $A$8 and penalties $A$9.

Each outer iteration consists of five parts: a diffusion denoising step, an $b$0-update, TV shrinkage updates for $b$1, a low-rank SVT update for $b$2, and multiplier updates. The $b$3-update solves

$b$4

For the TV subproblems, the shrinkage operator is

$b$5

equivalently

$b$6

For the low-rank term, the paper applies singular value thresholding (SVT): $b$7 Using the decomposition

$b$8

the singular values are shrunk according to

$b$9

The multiplier updates are

$$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$0

$$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$1

$$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$2

The paper states that the objective decreases monotonically and, under mild convexity assumptions, the method converges at rate $$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$3 (Deng et al., 5 Oct 2025). Since the full framework includes a diffusion prior, a plausible implication is that this convergence statement pertains to the ADMM-structured optimization as formulated in the paper rather than to a generic guarantee for all learned-prior inverse problems.

5. Computational design: PCG–FFT acceleration and slice-based 3D workflow

The dominant computational bottleneck is the $$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$4-subproblem, which involves the large linear system built from $$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$5, discrete gradient operators, and the identity term. TV-LoRA solves this system with PCG and an FFT-accelerated module. The stated reason is that $$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$6, $$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$7, and identity terms exhibit convolution-like structure that can be efficiently handled in the frequency domain, thereby reducing the per-iteration cost.

The acceleration claim is supported by ablation: disabling the PCG–FFT component does not materially change reconstruction quality, but increases runtime from 23 ms to 103 ms per iteration, corresponding to about a 4.5× speedup when the component is enabled (Deng et al., 5 Oct 2025). This makes the FFT-PCG module an efficiency mechanism rather than an accuracy mechanism.

Although the framework is aimed at 3D CT reconstruction, the implementation is explicitly slice-based. The 3D volume is processed as a sequence of 2D slices; projection data are generated from the 3D Radon transform; and reconstructions are evaluated in the axial, coronal, and sagittal views. The paper therefore distinguishes between the overall application target—efficient large-scale 3D reconstruction—and the actual computational strategy, which is 2D slice-oriented and GPU-parallel. A common misunderstanding would be to equate “3D CT reconstruction” with a fully volumetric 3D network or solver. The paper instead reports a slice-based workflow that supports 3D reconstruction through sequential or parallel slice processing.

6. Experimental protocol, quantitative results, and stated implications

The evaluation uses three public CT datasets: AAPM-2016 LDCT, CTHD, and LIDC-IDRI. For AAPM-2016 LDCT, the data are abdominal CT from 198 patients, with slices resized to $$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$8 and normalized to $$\min_x \frac{1}{2}\|Ax-b\|_2^2 + \lambda_{\mathrm{diff}}\bigl[-\log p_\theta(x)\bigr] + \alpha\,\mathrm{TV}(x) + \beta \|P(x)\|_* .$$9. CTHD is described as a liver lesion dataset with 3D volumes preprocessed to $\frac{1}{2}\|Ax-b\|_2^2$0. LIDC-IDRI contains thoracic CT scans for lung nodule analysis and is center-cropped and rescaled. Sparse-view reconstructions are tested at

$\frac{1}{2}\|Ax-b\|_2^2$1

with uniformly sampled projection angles

$\frac{1}{2}\|Ax-b\|_2^2$2

and detector size

$\frac{1}{2}\|Ax-b\|_2^2$3

Gaussian noise may be added in projection space to simulate low-dose conditions. The reported hardware and software stack includes an Intel Xeon 8358P CPU, 2 × NVIDIA RTX 4090 GPUs, Ubuntu 22.04, CUDA 11.8, cuDNN 8.7, and PyTorch 2.0.1.

The reported metrics are PSNR and SSIM across axial, coronal, and sagittal slices. The paper’s main quantitative claim is that TV-LoRA consistently outperforms the baselines across datasets and view counts, especially in SSIM, which is presented as reflecting structural fidelity. On LDCT, the best reported settings are approximately: for 8 views, axial $\frac{1}{2}\|Ax-b\|_2^2$4, coronal $\frac{1}{2}\|Ax-b\|_2^2$5, sagittal $\frac{1}{2}\|Ax-b\|_2^2$6; for 4 views, axial $\frac{1}{2}\|Ax-b\|_2^2$7, coronal $\frac{1}{2}\|Ax-b\|_2^2$8, sagittal $\frac{1}{2}\|Ax-b\|_2^2$9; for 2 views, axial $-\log p_\theta(x)$0, coronal $-\log p_\theta(x)$1, sagittal $-\log p_\theta(x)$2. On CTHD, the averages are around 33.67 dB / 0.861 at 8 views, 32.02 dB / 0.839 at 4 views, and 29.16 dB / 0.797 at 2 views; the paper states that TV-LoRA clearly beats ADMM-TV, TGV, FBPConvNet, and diffusion baselines in most settings. On LIDC, where absolute numbers are lower, TV-LoRA remains the best or near-best across views, with about 25.42 / 0.626 at 8 views, 21.67 / 0.545 at 4 views, and 18.68 / 0.465 at 2 views (Deng et al., 5 Oct 2025).

The qualitative findings emphasize preservation of fine soft-tissue boundaries, restoration of subtle textures, improved edge clarity, and suppression of streak artifacts and “plastic-like” oversmoothing, including stability under 2-view reconstruction. The baselines are reported to blur textures or leave pronounced streaking, especially under extreme sparsity.

The ablation study, conducted at $-\log p_\theta(x)$3, isolates the major components. Removing the LoRA / nuclear-norm term reduces SSIM by about 0.034, indicating the importance of the low-rank constraint for global texture continuity and coherent anatomical detail. Removing the diffusion prior causes a substantially larger degradation—about 2.27 dB PSNR and about 0.045 SSIM—and yields noticeably more streak artifacts. Removing PCG–FFT changes quality little but increases runtime from 23 ms to 103 ms per iteration. The paper therefore attributes the method’s performance to the complementarity of diffusion = realistic generative prior, low-rank LoRA = global structural regularization, TV = local edge preservation, and PCG–FFT = practical speed.

The intended clinical use is explicitly tied to extremely sparse-view and low-dose CT, particularly where acquisition dose must be minimized and only a very small number of projections are available. The listed scenarios are low-dose screening CT, sparse-sampling CT to reduce radiation, and reconstruction tasks in which retaining diagnostic detail is more important than simply denoising. This suggests that TV-LoRA is positioned not merely as an image enhancement method, but as a reconstruction framework for dose-constrained and sampling-constrained CT acquisition.