EPCT for Sparse-View CT Reconstruction

• EPCT is a training strategy that simulates multi-step artifact propagation in generalized diffusion-based CT reconstruction to correct structured errors.
• It leverages an EMA network to propagate reconstruction errors, exposing the restoration network to composite artifact states for enhanced robustness.
• EPCT demonstrates significant gains, with up to 3.8 dB PSNR improvement and sharper anatomical boundaries in sparse-view CT, as shown in ablation studies.

Searching arXiv for the specified papers to ground the article and verify citation details. Searching for (Chen et al., 14 Aug 2025) and related papers on Error-Propagating Composite Training. Error-Propagating Composite Training (EPCT) is a training strategy introduced in the Cross-view Generalized Diffusion Model (CvG-Diff) for sparse-view CT reconstruction. In that setting, EPCT explicitly simulates multi-step artifact propagation under generalized diffusion and trains the reconstruction network to correct the resulting states, rather than training only on single-level degraded-to-clean pairs. The method addresses a specific failure mode of generalized diffusion in sparse-view CT: intermediate reconstruction errors are structured streaks and blurring, not Gaussian perturbations, and therefore can be redistributed and amplified across reverse steps unless the network is exposed to such states during training (Chen et al., 14 Aug 2025).

1. Placement within generalized diffusion for sparse-view CT

Sparse-view CT reconstruction seeks to recover a high-quality CT image $x_0$ from very few projection views such as $N_v = 18, 36, 72$. Classical filtered back-projection (FBP) under these conditions produces severe streak artifacts. Single-step deep networks can suppress such artifacts, but under significant sparsity they tend to over-smooth and lose fine structures. Diffusion models improve reconstruction through iterative refinement and generative priors, yet classical denoising diffusion models assume Gaussian degradation, require hundreds or thousands of sampling steps, and become unstable in ultra-sparse regimes (Chen et al., 14 Aug 2025).

CvG-Diff reformulates sparse-view CT reconstruction as a generalized diffusion process, following Cold Diffusion. Its degradation operator is deterministic and encodes CT artifact formation directly: $$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$ where $\mathcal{A}$ is the Radon transform, $\mathcal{P}(T)$ applies an angular subsampling mask at severity level $T$, and $\mathcal{A}^\dagger$ denotes FBP back to the image domain. Different levels $t$ correspond to different view counts $\mathcal{T}_t$ via a mapping $g(t)$, enabling cross-view correlations: higher-view reconstructions are less corrupted and share semantic structure and radial artifact patterns with lower-view ones (Chen et al., 14 Aug 2025).

Within this framework, the restoration network $N_v = 18, 36, 72$0 learns to map a degraded image at level $N_v = 18, 36, 72$1 to the clean target $N_v = 18, 36, 72$2 with

$N_v = 18, 36, 72$3

Sampling iteratively applies restoration and degradation adjustment: $N_v = 18, 36, 72$4 The central issue is that, in sparse-view CT, errors in $N_v = 18, 36, 72$5 are structured artifacts. A model trained only on $N_v = 18, 36, 72$6 never observes the propagated artifact patterns created by this multi-step update. EPCT is the mechanism designed to fill that gap (Chen et al., 14 Aug 2025).

2. Formal definition of EPCT

EPCT is a two-part training strategy for the reconstruction network $N_v = 18, 36, 72$7. The first part is standard restoration training at arbitrary severity levels. The second part, error-propagating composite training, uses an EMA version of the reconstruction network to generate an intermediate reconstruction at a chosen level $N_v = 18, 36, 72$8, propagates that reconstruction to a lower level $N_v = 18, 36, 72$9 with the generalized diffusion update rule, and then trains the current network to reconstruct the clean image from the resulting propagated-artifact state (Chen et al., 14 Aug 2025).

The additional EPCT objective is

$$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$0

where $$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$1 is not simply $$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$2, but a state synthesized by transporting reconstruction error from level $$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$3 to level $$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$4. The training-time EPCT construction proceeds as follows. First, sample a target severity level $$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$5 and compute

$$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$6

Then sample a lower intermediate level $$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$7. An EMA network $$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$8 produces a stable approximation

$$x_T = D(x_0,T) = \mathcal{A}^\dagger \mathcal{P}(T)\mathcal{A}x_0,$$9

The propagated state is then built with

$\mathcal{A}$0

Finally, the current network is optimized to recover $\mathcal{A}$1 from $\mathcal{A}$2 through $\mathcal{A}$3 (Chen et al., 14 Aug 2025).

The EMA copy is updated by

$\mathcal{A}$4

with $\mathcal{A}$5 and update interval $\mathcal{A}$6. No explicit weighting coefficients between restoration and composite losses are specified, and neither masks nor region-wise weights are introduced explicitly in the equations. The method instead relies on the fact that propagated states naturally concentrate errors in locations where prior reconstructions failed (Chen et al., 14 Aug 2025).

3. Mechanism of propagated-artifact correction

EPCT is motivated by the structure of error transport under sparse-view CT physics. When a reconstruction $\mathcal{A}$7 contains mistakes, the forward projection $\mathcal{A}$8 converts them into sinogram inconsistencies, angular subsampling $\mathcal{A}$9 filters those inconsistent projections, and FBP $\mathcal{P}(T)$0 maps them back into the image domain as streak artifacts and blurred structures. The reverse update

$\mathcal{P}(T)$1

therefore transports structured errors from one level to the next rather than behaving like stochastic denoising (Chen et al., 14 Aug 2025).

EPCT reuses this same mechanism during training. The synthesized input

$\mathcal{P}(T)$2

contains the mismatch $\mathcal{P}(T)$3, which is large exactly where the EMA reconstruction fails to match the degraded image at level $\mathcal{P}(T)$4. Those regions correspond to streaks, blurring, or hallucinated structure. Re-degradation to level $\mathcal{P}(T)$5 then reprojects these mismatches through the CT operators, generating artifact patterns at the new severity level. Training on such states is what allows the model to “identify error-prone regions for correction,” even though no explicit attention or masking module is introduced (Chen et al., 14 Aug 2025).

This procedure improves robustness to large inter-step jumps. Sampling across view counts such as 18, 36, and 72 with only a few generalized diffusion steps is efficient, but it is also the regime in which propagated errors can grow rapidly. EPCT teaches the model to invert not only the physical degradation $\mathcal{P}(T)$6, but the composition of physical degradation and propagated reconstruction error. A plausible implication is that EPCT is best understood as distribution matching between training-time inputs and the non-ideal intermediate states encountered at inference time.

4. Role inside the CvG-Diff pipeline

EPCT is a training-time component; Semantic-Prioritized Dual-Phase Sampling (SPDPS) is an inference-time component. Their functions are complementary. EPCT makes the restoration network robust to propagated artifacts and large inter-step jumps. SPDPS then allocates a small sampling budget across semantic correction steps at sparse-view levels and detail refinement steps at denser-view levels, with the aim of achieving stable few-step reconstruction (Chen et al., 14 Aug 2025).

In the overall CvG-Diff pipeline, training uses the deterministic degradation operator $\mathcal{P}(T)$7 together with the cross-view mapping

$\mathcal{P}(T)$8

Inference starts from the FBP reconstruction at the target sparsity, then applies SPDPS with $\mathcal{P}(T)$9, $T$0, and $T$1, so that $T$2 steps are allocated to semantic correction. SPDPS monitors structural convergence via SSIM between consecutive reconstructions and, when the SSIM threshold is exceeded, resets the degradation level back to the input sparse-view level through

$T$3

This reset uses improved intermediate reconstructions to better identify and correct error-prone regions at the original sparsity level, where artifacts are strongest (Chen et al., 14 Aug 2025).

With EPCT and SPDPS together, CvG-Diff achieves 38.34 dB PSNR and 0.9518 SSIM for 18-view CT using only 10 steps on the AAPM-LDCT dataset. The paper also states that extensive experiments demonstrate superiority over state-of-the-art sparse-view CT reconstruction methods. Within that result, EPCT is the component that stabilizes generalized diffusion sufficiently for few-step SPDPS sampling to be effective (Chen et al., 14 Aug 2025).

5. Empirical effects, implementation profile, and limits

The empirical effect most directly attributed to EPCT appears in the ablation study. Adding EPCT without SPDPS increases performance from 33.67 dB / 0.8242 SSIM to 37.85 dB / 0.9468 SSIM at 18 views; from 37.02 dB / 0.8915 SSIM to 41.48 dB / 0.9687 SSIM at 36 views; and from 43.38 dB / 0.9730 SSIM to 45.66 dB / 0.9854 SSIM at 72 views. The paper describes this as an approximately 3.8 dB average PSNR gain and attributes it directly to EPCT, interpreting the result as evidence that training on propagated-artifact states improves both stability and reconstruction quality (Chen et al., 14 Aug 2025).

Qualitative evidence is described in Fig. 3 and Fig. 4 of the paper. Without EPCT, multi-step reconstruction yields persistent streaks and blurred boundaries, and error maps show large deviations around high-contrast edges and fine structures. With EPCT, those regions are better corrected, leading to sharper anatomical boundaries and lower error magnitudes. This suggests that the method’s practical benefit is not limited to global fidelity metrics, but extends to anatomically salient structures (Chen et al., 14 Aug 2025).

The implementation details define the scope in which these results were obtained. The reconstruction network is a Diffusion UNet with residual blocks, base feature dimension 128, and channel multipliers $T$4. Training uses the AAPM-LDCT dataset with 5,936 CT slices from 10 patients, split into 9 patients for train/validation (5,410 slices) and 1 patient for test (526 slices). Sparse-view simulation uses fan-beam geometry, source–detector distance 59.5 cm, 672 detector elements, 120 kVp, and 500 mA, implemented with TorchRadon. Optimization runs for 40 epochs with batch size 4, Adam with $T$5, initial learning rate $T$6, and decay by 0.8 after epoch 25. The paper does not explicitly list failure cases for EPCT. Indirectly, it notes that EPCT alone still benefits from adaptive sampling, and that EPCT assumes stable EMA reconstructions; if the EMA model fails badly early in training, the propagated states could be poor, although EMA smoothing and gradual training are said to mitigate this (Chen et al., 14 Aug 2025).

6. Broader conceptual uses and terminological scope

The phrase “Error-Propagating Composite Training” also appears in broader explanatory usage outside CvG-Diff, but not always as a paper’s formal method name. In the CPR-Coach / ImagineNet setting, it is used to describe a regime in which the model learns representations of primitive error actions from single-error supervision and then propagates those representations to recognize composite error actions that were never seen during training (Wang et al., 2023). There, the relevant mechanism is feature-level composition: $T$7 paired with multi-hot labels

$T$8

and binary cross-entropy over 14 classes. The task setting is Single-class Training & Multi-class Testing, with 13 single-error classes plus 1 Correct class in Set-1 and 74 composite error classes in Set-2. This use of EPCT refers to propagating single-error knowledge into composite feature space rather than to artifact transport in generalized diffusion (Wang et al., 2023).

A second conceptual extension appears in “Composite Optimization with Error Feedback: the Dual Averaging Approach,” where the phrase is used to motivate training or optimization with compressed gradients, error feedback, and a composite objective $T$9 (Gao et al., 3 Oct 2025). In that context, the relevant propagated quantity is the compression residual

$\mathcal{A}^\dagger$0

which is accumulated in a Dual Averaging framework with EControl. The paper’s central claim is that standard Error Feedback methods fail in the broader setting of composite optimization, and that Dual Averaging combined with EControl yields, for the first time, a strong convergence analysis for composite optimization with error feedback (Gao et al., 3 Oct 2025).

This suggests that EPCT is not yet a single standardized term across arXiv. In the strict sense, it denotes the CvG-Diff training strategy for sparse-view CT reconstruction (Chen et al., 14 Aug 2025). In broader interpretive use, it names a recurring design pattern in which structured errors or residuals are deliberately propagated into training or optimization so that a model learns to handle the composite states that arise at deployment time (Wang et al., 2023, Gao et al., 3 Oct 2025).