DM4CT: Diffusion CT Methods & Benchmarks

• DM4CT is a polysemous research label for CT, covering diffusion-based reconstruction, mutual-domain material decomposition, and MRI-to-CT synthesis.
• Diffusion methods in DM4CT excel in challenging, sparse-view and noisy conditions by balancing learned priors with data consistency, yet they face issues like artifact suppression and computational demands.
• The benchmark integrates diverse datasets and evaluation protocols, offering insights into performance trade-offs and guiding improvements in CT reconstruction and harmonization techniques.

DM4CT denotes several distinct constructs in recent computed-tomography research. Its most explicit and broadest usage is as the name of a benchmark for diffusion-based CT reconstruction, designed to compare diffusion priors with classical, model-based, unsupervised, and supervised reconstruction pipelines across medical, industrial, and real-world synchrotron settings (Shi et al., 20 Feb 2026). In adjacent literature, the same label has also been used for mutual-domain material decomposition in dual-energy CT, for a conditional 3D denoising diffusion model for MRI-to-CT synthesis, and for CT standardization and enhancement in a learned latent space (Su et al., 2020, Pan et al., 2023, Selim et al., 2023). As a result, DM4CT is best understood as a polysemous research label rather than a single universally standardized method.

1. Terminology and scope

In the literature summarized here, DM4CT refers to multiple CT-centered formulations. The 2026 benchmark paper makes DM4CT a dataset-and-evaluation framework for reconstruction, whereas earlier uses attach the label to specific architectures or application pipelines.

Usage of DM4CT	Task	Paper
Benchmark for CT reconstruction	Diffusion-method evaluation across datasets and regimes	(Shi et al., 20 Feb 2026)
Mutual-Domain Material Decomposition for CT	Dual-energy CT material decomposition via DIRECT-Net	(Su et al., 2020)
Conditional 3D denoising diffusion model	MRI-to-CT synthetic CT generation	(Pan et al., 2023)
CT standardization and enhancement	Latent-space diffusion harmonization	(Selim et al., 2023)

This multiplicity matters methodologically. Some DM4CT usages concern inverse problems with explicit forward operators, some concern cross-modality synthesis, and some concern harmonization of already reconstructed images. A plausible implication is that comparisons between papers using the same acronym require attention to the underlying task definition, observation model, and evaluation protocol.

2. DM4CT as a reconstruction benchmark

As a benchmark, DM4CT formalizes CT reconstruction as recovery of an unknown volume $x \in \mathbb{R}^m$ from projections $b \in \mathbb{R}^n$, with the nominal linear model $b = A x$, while emphasizing that practical CT deviates from the idealized setting through correlated, heteroscedastic noise, structured artifacts, geometry dependencies, and value-range misalignment (Shi et al., 20 Feb 2026). The benchmark explicitly notes that after the log-transform, noise variance becomes intensity-dependent, with

$$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$

and that ring artifacts can be modeled by corrupting a fraction $p_{\rm ring}$ of detector columns via

$y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$

The reconstruction objective is written in regularized form as

$$\min_{x \in \mathbb{R}^m} \; \|A x - b\|^2 + \lambda \,\mathcal{R}(x),$$

with diffusion-based approaches interpreting $\mathcal{R}$ implicitly through the posterior

$p(x \mid b) \propto p(x)\,p(b \mid x).$

The benchmark describes reverse-time conditioning through a diffusion SDE,

$$d x_t = \bigl[ -\tfrac12 \beta_t x_t - \beta_t(\nabla_{x_t}\log p(x_t)+\nabla_{x_t}\log p(b\mid x_t)) \bigr]dt + \sqrt{\beta_t}\,d\bar w_t,$$

where the score network approximates $b \in \mathbb{R}^n$0, and data consistency approximates $b \in \mathbb{R}^n$1 (Shi et al., 20 Feb 2026).

The benchmark organizes data-consistency strategies into gradient steering, optimization steps, pseudoinverse guidance, and variational Bayes. This framing is significant because it shifts the evaluation focus away from diffusion backbones alone and toward the coupled prior-likelihood mechanism. In CT, the dominant issue is not merely generative fidelity, but how the prior interacts with geometry, projection sparsity, and non-ideal acquisition statistics.

3. Dataset design, methods, and evaluation protocol

DM4CT comprises three principal datasets: medical CT from the AAPM Low-Dose CT Grand Challenge 2016, industrial CT from LoDoInd, and a real-world synchrotron CT dataset acquired at 24 keV in parallel-beam geometry (Shi et al., 20 Feb 2026). The medical dataset contains volumes ranging from $b \in \mathbb{R}^n$2 to $b \in \mathbb{R}^n$3, with nine volumes for training and one for testing. The industrial dataset uses central slices from a $b \in \mathbb{R}^n$4 volume, with 3,000 slices for training and 500 for testing. The synchrotron dataset uses two rock samples, with one for training and one for testing, and evaluates reconstructions from 200, 100, and 60 evenly subsampled projections.

For the medical and industrial settings, five simulation configurations are defined: 40 angles without noise; 20 angles with Poisson noise at $b \in \mathbb{R}^n$5; 80 angles with stronger noise at $b \in \mathbb{R}^n$6; 80 angles with ring artifacts using $b \in \mathbb{R}^n$7 and $b \in \mathbb{R}^n$8; and 40 angles over $b \in \mathbb{R}^n$9 without added noise (Shi et al., 20 Feb 2026). The real-world synchrotron configuration omits ring correction, which is methodologically important because it exposes reconstruction methods to non-simulated artifact structure.

The benchmark evaluates ten diffusion-based methods: MCG, DPS, PSLD, PGDM, DDS, ReSample, DMPlug, Reddiff, HybridReg, and DiffStateGrad. It also compares against FBP, SIRT, ADMM-PDTV, FISTA-SBTV, DIP, INR, R²-Gaussian, and SwinIR (Shi et al., 20 Feb 2026). The abstract characterizes these as “seven strong baselines,” while the detailed method list enumerates eight named baseline families or methods; this discrepancy is present in the source summary and is therefore part of the record rather than an external correction.

Quantitative evaluation includes MSE, PSNR, SSIM, RMSE, LPIPS, and the data-fit loss $b = A x$0, together with inference time and GPU memory for single-slice reconstruction under a 40 GB limit, and training-time resource measurements for diffusion backbones and SwinIR (Shi et al., 20 Feb 2026). Hyperparameters are tuned by grid search on held-out training slices by minimizing mean-squared error. This protocol gives DM4CT a dual role: it is not only a ranking benchmark, but also a stress test for the practical deployability of diffusion methods under realistic resource constraints.

4. Empirical findings and methodological interpretation

The central empirical result is that diffusion methods generally surpass classical reconstruction methods such as FBP and SIRT, and model-based TV variants such as ADMM-PDTV and FISTA-SBTV, under sparse-view and noisy conditions, yet still underperform the supervised SwinIR baseline overall (Shi et al., 20 Feb 2026). At the same time, INR remains competitive in noise-free scenarios and on the real-world synchrotron dataset. No single diffusion method dominates across all configurations.

Within the diffusion cohort, ReSample and DPS frequently rank near the top, while PGDM and MCG perform strongly on data-fit metrics. DDS is reported to suffer under Poisson noise because of its Gaussian likelihood assumption (Shi et al., 20 Feb 2026). These results underscore that performance depends not only on score quality but on the compatibility between the assumed likelihood correction and the measurement statistics.

Qualitative analysis adds a more nuanced picture. Diffusion reconstructions often recover fine texture and small structures “realistically,” but may diverge from the true anatomy through hallucinations, which can lower PSNR and SSIM despite visually plausible appearance (Shi et al., 20 Feb 2026). By contrast, DIP and INR are described as overly smooth, with loss of high-frequency detail, yet they may align better with MSE-oriented metrics in clean settings. This tension illustrates a recurrent inverse-problem trade-off between perceptual realism and strict image fidelity.

DM4CT also analyzes posterior uncertainty by sampling MCG ten times on the same measurement; the resulting standard deviations peak at object edges and ambiguous regions (Shi et al., 20 Feb 2026). Range-null-space decomposition further differentiates methods: ReSample, with strict consistency steps, exhibits low null-space energy and strong data fidelity but may overfit noisy artifacts, whereas gradient-based methods such as DPS admit higher null-space content and therefore stronger prior influence. Latent diffusion brings additional difficulties: PSLD can produce blocky discontinuities when enforcing gradients in latent space, while ReSample’s explicit optimization mitigates them in noise-free settings but can overfit in the presence of noise. Collectively, these observations make the benchmark less a single leaderboard than a diagnostic map of prior-data coupling regimes.

5. Related task-specific DM4CT formulations and adjacent diffusion pipelines

Beyond the benchmark, DM4CT has been used for domain-specific CT pipelines. In dual-energy CT, the DIRECT-Net paper is summarized as a “DM4CT” formulation meaning mutual-Domain Material Decomposition for CT (Su et al., 2020). There, the observation model begins from two spectra $b = A x$1 and $b = A x$2, with attenuation expanded in two basis materials,

$b = A x$3

DIRECT-Net learns an end-to-end map from paired sinograms $b = A x$4 to basis maps $b = A x$5 by coupling an 8-layer Sinogram-Domain SubNet, a differentiable FBP-based Domain-Transform Module, and an Image-Domain SubNet. On the XCAT water/bone task, it reports ROI means and standard deviations closer to truth than several comparators; on the iodine phantom, concentration error is reported as $b = A x$6, with a linear-fit slope of approximately $b = A x$7 and intercept approximately $b = A x$8 mg/cc; and inference time is listed as $b = A x$9 s versus approximately $$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$0–$$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$1 s for iterative baselines (Su et al., 2020). This usage is not diffusion-based, but it establishes an early “multi-domain” interpretation of DM4CT.

A separate usage appears in MRI-to-CT synthesis. The 3D transformer-based denoising diffusion model summarized as DM4CT defines a conditional reverse process $$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$2, where $$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$3 is a co-registered MRI volume and the backbone is a Swin-Vnet combining convolutional blocks with Swin attention (Pan et al., 2023). The model uses $$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$4 steps with $$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$5, performs inference on a resampled set of 50 timesteps, and averages five Monte-Carlo runs. Reported test metrics are MAE $$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$6 HU, PSNR $$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$7 dB, SSIM $$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$8, and NCC $$\mathrm{Var}[-\log(\hat I/I_0)] \approx e^{(A x)}/I_0,$$9 on the brain dataset, with corresponding prostate results of MAE $p_{\rm ring}$0 HU, PSNR $p_{\rm ring}$1 dB, SSIM $p_{\rm ring}$2, and NCC $p_{\rm ring}$3 (Pan et al., 2023). This formulation treats DM4CT as conditional synthesis rather than reconstruction from projections.

The latent-standardization line represented by DiffusionCT uses a U-Net-based encoder-decoder with a DDPM at the bottleneck to map non-standard CT images into a standardized domain (Selim et al., 2023). Training proceeds in two stages: encoder-decoder pretraining via $p_{\rm ring}$4 reconstruction, followed by latent DDPM training with a linear $p_{\rm ring}$5 schedule in $p_{\rm ring}$6 over $p_{\rm ring}$7 and an $p_{\rm ring}$8-prediction objective. On Siemens Br40$p_{\rm ring}$9Bl64 standardization, DiffusionCT is reported to preserve 64% more features at the 15% error threshold than raw inputs, outperforming GANai, STAN-CT, and RadiomicGAN under that criterion; inference is approximately 30 s per $y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$0 slice with 1,000 sampling steps (Selim et al., 2023). Here DM4CT denotes harmonization and enhancement, especially for radiomic reproducibility.

A broader diffusion-CT context is provided by two additional works that are not explicitly labeled DM4CT in their titles but are methodologically adjacent. One is a motion-corrected sparse-view 4DCT framework that combines a wavelet diffusion model with alternating motion-field updates, denoted JRM-Diff, and reports $y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$1 dB PSNR and $y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$2 SSIM on end-inhale XCAT reconstructions, versus $y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$3 dB and $y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$4 for JRM-TV (Paepe et al., 21 Jan 2025). The other is a conditional DDPM for FDCT-to-MDCT enhancement in thrombectomy workflow support, using $y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$5, a linear $y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$6 schedule from approximately $y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$7 to approximately $y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$8, a 2D U-Net-style encoder-decoder with a 64-head self-attention block at $y = y_0 + M \cdot \mathcal{N}(0,\sigma^2 I).$9 resolution, and 125 training volumes from the SPINNERS trial; it reports MSE $$\min_{x \in \mathbb{R}^m} \; \|A x - b\|^2 + \lambda \,\mathcal{R}(x),$$0 HU$$\min_{x \in \mathbb{R}^m} \; \|A x - b\|^2 + \lambda \,\mathcal{R}(x),$$1, SSIM $$\min_{x \in \mathbb{R}^m} \; \|A x - b\|^2 + \lambda \,\mathcal{R}(x),$$2, and PSNR $$\min_{x \in \mathbb{R}^m} \; \|A x - b\|^2 + \lambda \,\mathcal{R}(x),$$3 dB versus a pix2pix baseline at $$\min_{x \in \mathbb{R}^m} \; \|A x - b\|^2 + \lambda \,\mathcal{R}(x),$$4, $$\min_{x \in \mathbb{R}^m} \; \|A x - b\|^2 + \lambda \,\mathcal{R}(x),$$5, and $$\min_{x \in \mathbb{R}^m} \; \|A x - b\|^2 + \lambda \,\mathcal{R}(x),$$6 dB, respectively (Corbaz et al., 22 Aug 2025). These works suggest that the benchmarked issues in DM4CT—data consistency, artifact suppression, and failure modes under degraded input—recur across multiple CT-adjacent diffusion tasks.

6. Limitations, open problems, and research significance

DM4CT’s main contribution is not the proposal of a single reconstruction algorithm but a systematic characterization of when diffusion priors help and when they fail (Shi et al., 20 Feb 2026). The benchmark concludes that diffusion priors excel under severe ill-posedness, especially sparse angles and high noise, because they inject strong learned structure. It also concludes that balancing prior strength against data consistency is critical: insufficient steering leads to prior-dominated hallucinations, whereas overly aggressive steering produces denoising collapse or noise overfitting.

The real-world synchrotron experiments expose the current limits of diffusion CT. Domain shift, limited training data, misaligned value ranges, and unmodeled noise characteristics degrade performance relative to simulated settings (Shi et al., 20 Feb 2026). This observation aligns with failure modes reported in adjacent application papers. In FDCT enhancement, extremely strong motion or beam-hardening can induce hallucinated structures or lesion suppression, including one case where a small intraparenchymal bleed was erased (Corbaz et al., 22 Aug 2025). In MRI-to-CT synthesis and latent CT standardization, inference remains substantially slower than GAN-based alternatives because sampling still requires tens to thousands of denoising steps (Pan et al., 2023, Selim et al., 2023). A plausible implication is that CT diffusion research remains bottlenecked as much by likelihood modeling, calibration, and acceleration as by score-network capacity.

The forward-looking agenda is correspondingly concrete. The benchmark highlights better noise models, including Poisson-aware conditioning and data-dependent covariance; combinations of diffusion with INRs; flow-based priors such as FlowDPS or other hybrid generative models; extension to multi-institutional and multi-geometry datasets; and systematic downstream evaluation through tasks such as organ segmentation and radiologist scoring (Shi et al., 20 Feb 2026). In the related task-specific literature, natural extensions include full 3D DDPMs for slice-wise pipelines, uncertainty quantification for hallucination detection, and improved retention of pathology under severe artifact regimes (Corbaz et al., 22 Aug 2025).

Taken together, DM4CT marks a transition in CT methodology from isolated demonstrations of diffusion priors toward a comparative science of inverse-problem solvers. The term now spans benchmarking, decomposition, synthesis, and harmonization. Its unifying theme is the attempt to integrate learned generative structure with the physical and statistical constraints of X-ray imaging.