Semantic-Prioritized Dual-Phase Sampling

• SPDPS is an inference-time sampling strategy that reorganizes the generalized diffusion process into a semantic correction phase followed by detail refinement for sparse-view CT.
• It employs an adaptive reset mechanism based on SSIM to correct structural errors early, mitigating artifact propagation and improving anatomical accuracy.
• The subsequent detail-refinement phase sharpens edges and eliminates residual artifacts, achieving high-quality reconstructions with efficient network evaluations.

Semantic-Prioritized Dual-Phase Sampling (SPDPS) is the inference-time sampling strategy introduced in the Cross-view Generalized Diffusion Model (CvG-Diff) for sparse-view computed tomography (CT) reconstruction. It reformulates the reverse trajectory of generalized diffusion so that semantic correctness is prioritized before detail refinement: the first phase remains close to the sparse-view regime and uses adaptive resetting to correct structural and anatomical errors, while the second phase proceeds through denser-view levels to sharpen edges and remove residual artifacts. Within CvG-Diff, SPDPS is intended to mitigate artifact propagation and the inefficiency of strictly sequential generalized diffusion, and the reported result on AAPM-LDCT for 18-view CT is 38.34 dB PSNR and 0.9518 SSIM using 10 steps (Chen et al., 14 Aug 2025).

1. Position within generalized diffusion for sparse-view CT

CvG-Diff adopts generalized diffusion, or cold diffusion, rather than standard DDPM-type stochastic diffusion. In this formulation, a degradation operator $D$ maps a clean image $x_0$ to a degraded image $x_t$ at severity level $t$, and a restoration network $R_\theta$ approximates the inverse. The training objective is

$$\mathcal{L}_{\text{restore}} = \left\| R_\theta(D(x_0, t), t) - x_0 \right\|_2 .$$

The corresponding reverse update is

$$\hat{x}_0^t = R_\theta(x_t, t), \qquad x_{t-1} = x_t - D(\hat{x}_0^t, t) + D(\hat{x}_0^t, t-1),$$

and sequential application over timesteps $t \in \{T, T-1, \dots, 1\}$ is denoted by $I(x_T, T)$ (Chen et al., 14 Aug 2025).

For sparse-view CT, CvG-Diff replaces stochastic Gaussian degradation with deterministic artifact generation:

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

where $x_0$0 is the Radon transform, $x_0$1 is the angular subsampling operator in sinogram space, and $x_0$2 is filtered backprojection (FBP). Severity level $x_0$3 is therefore tied to a view count $x_0$4, ranging from very sparse settings such as 18 views to denser settings such as 288 views. The same restoration model is trained across these severity levels.

The central problem addressed by SPDPS arises from direct use of the standard generalized diffusion update. If an intermediate reconstruction $x_0$5 contains errors, then reapplying $x_0$6 projects those errors into sinogram space and back, producing additional streak artifacts. The reported consequence is artifact propagation: early errors from highly sparse inputs produce blurred or incorrect anatomy, and later denser-view updates mainly refine high-frequency detail rather than fully correcting those semantic errors.

2. Semantic-prioritized phase

The first phase of SPDPS is explicitly semantic-prioritized. In the terminology used for CvG-Diff, semantic correctness denotes the global, clinically relevant structure: organ shapes and boundaries, bone contours, lesion presence and approximate extent, and the overall anatomical layout without gross distortions or major blurring. Detail refinement, by contrast, concerns fine edges, texture, soft-tissue contrast, minor streaks, ringing, and small intensity errors. The design premise is that in extremely sparse settings, getting the anatomy right first is more important than immediately refining low-level detail (Chen et al., 14 Aug 2025).

Inference starts from the sparse-view FBP image $x_0$7. SPDPS then follows generalized diffusion from level $x_0$8 downward, but augments the trajectory with an adaptive reset rule. At each transition from $x_0$9 to $x_t$0, it computes the similarity

$x_t$1

where $x_t$2 and $x_t$3. If

$x_t$4

the model treats this as anatomical convergence between consecutive severity levels. At that point, SPDPS resets the degradation level back near the sparse-view input:

$x_t$5

This reset combines three elements: the original sparse-view artifact structure $x_t$6, the improved reconstruction $x_t$7, and a slightly denser-view degradation of that improved reconstruction. The intended effect is to expose error-prone regions while preserving the semantic gains already obtained. From the new state $x_t$8, generalized diffusion resumes at level $x_t$9, and the reset-and-refine cycle repeats until the semantic-prioritized budget is exhausted. The paper’s interpretation is that this procedure prevents the sampler from spending iterations over-refining anatomically stable regions while major structural errors persist.

3. Detail-refinement phase and phase transition

After the semantic-prioritized budget has been used, SPDPS switches to a detail-refinement phase. At that stage, the intermediate image is assumed to be anatomically stable and to lie at a relatively denser-view severity level. The sampler then follows standard generalized diffusion without further resets:

$t$0

that is, $t$1 reverse steps in the denser-view regime. In expanded form,

$t$2

for decreasing severity levels $t$3. The operational distinction from the first phase is exact: there are no SSIM checks and no reset back to the original sparse-view level; the trajectory becomes monotone toward less severe degradation (Chen et al., 14 Aug 2025).

The transition is governed by a fixed allocation of steps together with adaptive resets inside the semantic portion. For the AAPM-LDCT experiments, the reported configuration is total NFE $t$4, detail-refinement steps $t$5, semantic-prioritized steps $t$6, and SSIM threshold $t$7. The same diffusion UNet is used in both phases. SPDPS therefore does not introduce a separate architecture for semantics and detail; it changes the trajectory in severity-space rather than the network structure itself.

Several potential misconceptions are explicitly resolved by this formulation. SPDPS does not define a separate semantic loss; semantic prioritization is realized through phase partitioning, the SSIM-based convergence criterion, and the reset operation. It also does not directly use multiple distinct scan inputs. Instead, it implicitly exploits cross-view structure through the severity-dependent degradation operator and the fact that the restoration network is trained across multiple sparsity levels.

4. Interaction with EPCT and cross-view correlations

SPDPS is paired with Error-Propagating Composite Training (EPCT), which is a training-time strategy rather than an inference-time scheduler. EPCT is designed to expose the restoration network to the kinds of propagated-artifact states that arise during multi-step generalized diffusion. It maintains an EMA network,

$t$8

and constructs composite training samples through

$t$9

with composite loss

$R_\theta$0

This forces the model to correct propagated artifacts at intermediate levels rather than only denoising or de-artifacting one step at a time (Chen et al., 14 Aug 2025).

The relationship between EPCT and SPDPS is functional. SPDPS relies on the model being robust to large inter-level jumps and reset states such as $R_\theta$1; EPCT supplies that robustness during training. The paper states that there is no explicit EPCT-derived error map passed to SPDPS. Instead, EPCT shapes network behavior so that the adaptive schedule becomes effective at test time.

Cross-view correlations are also mediated through the degradation model and the training regime. The deterministic operator $R_\theta$2 maps severity level to view count, EPCT samples multiple target and intermediate levels, and SPDPS repeatedly resets from $R_\theta$3 to $R_\theta$4 while using reconstructions informed by multiple sparsity levels. In that sense, SPDPS leverages cross-view structure implicitly rather than through separate measurements from multiple acquisitions.

5. Quantitative performance, ablations, and parameter sensitivity

The reported evidence positions SPDPS as a contributor to both reconstruction quality and sampling efficiency. On AAPM-LDCT, CvG-Diff with EPCT and SPDPS achieves the following results in 10 network function evaluations (Chen et al., 14 Aug 2025):

View setting	PSNR	SSIM
18-view	38.34 dB	95.18%
36-view	41.78 dB	97.05%
72-view	45.94 dB	98.63%

The same report gives an inference time of 0.68 s. For comparison, standard diffusion VSS with NFE=1000 is reported at 32.34 dB and 87.90% SSIM for 18-view CT, with time 264.71 s, while CoSIGN with NFE=10 is reported at 31.84 dB and 86.31% SSIM. The stated conclusion is that the combination of EPCT and SPDPS is especially advantageous in the extreme sparsity regime.

The ablation study isolates the effect of SPDPS relative to sequential sampling:

Variant	PSNR	SSIM
No EPCT, no SPDPS	33.67	82.42
SPDPS only	34.67	85.23
EPCT only	37.85	94.68
EPCT + SPDPS	38.34	95.18

These numbers show that EPCT accounts for the larger share of the gain, while SPDPS provides an additional improvement over EPCT with sequential sampling of approximately 0.5 dB PSNR and 0.5% SSIM for 18-view CT. The qualitative description in the paper attributes that margin particularly to correction of blurred anatomical boundaries.

Sensitivity analysis further characterizes the scheduler. For 18-view CT, $R_\theta$5 gives 38.34 dB and 95.18% SSIM; $R_\theta$6 gives 38.06 and 95.03%; and $R_\theta$7 gives 37.94 and 94.96%. For the number of detail steps, $R_\theta$8 gives 38.33 and 95.13%, $R_\theta$9 gives 38.34 and 95.15%, and $$\mathcal{L}_{\text{restore}} = \left\| R_\theta(D(x_0, t), t) - x_0 \right\|_2 .$$0 gives 38.34 and 95.18%. The reported interpretation is that performance is relatively robust to $$\mathcal{L}_{\text{restore}} = \left\| R_\theta(D(x_0, t), t) - x_0 \right\|_2 .$$1 so long as the semantic phase remains sufficiently strong, while overly high $$\mathcal{L}_{\text{restore}} = \left\| R_\theta(D(x_0, t), t) - x_0 \right\|_2 .$$2 delays resets and slightly reduces benefit.

6. Conceptual relation to dynamic dual sampling and broader scope

An interpretive lens provided alongside the sparse-view CT formulation relates SPDPS to the earlier Dynamic Dual Sampling Module (DDSM) for fine-grained semantic segmentation. DDSM is described as a semantic-guided, two-phase dynamic sampling and propagation mechanism between high-level and low-level features, with a spatial-wise phase over pixels and a channel-wise phase over channels. In that earlier setting, the module propagates high-level semantics to low-level detail through dynamic affinity modeling and adaptive compact support, yielding fine-grained segmentation with well-preserved boundaries (Shi et al., 2021).

This comparison does not identify DDSM as the same method as SPDPS. Rather, it suggests a shared design intuition: semantics are treated as the primary corrective signal, and fine detail is refined afterward or in a complementary second phase. The mechanisms differ substantially. DDSM is a plug-in module inserted between adjacent backbone stages in segmentation networks, whereas SPDPS is an inference-time schedule for generalized diffusion in sparse-view CT. Even so, both formulations exemplify a two-phase organization in which semantically informative guidance is privileged before or alongside detailed reconstruction.

The scope and limitations of SPDPS remain tied to its reconstruction setting. The reset mechanism depends on sparse-view CT degradation modeled by the deterministic Radon-transform, angular-subsampling, and FBP pipeline. Performance depends on suitable tuning of $$\mathcal{L}_{\text{restore}} = \left\| R_\theta(D(x_0, t), t) - x_0 \right\|_2 .$$3 and on EPCT training; if the underlying restoration model fails to learn correct semantics under propagated artifacts, SPDPS cannot fully compensate. The paper also notes that extremely pathological anatomy or unusual artifacts might challenge SSIM-based convergence detection. At the same time, the authors indicate that the broader principles—dual-phase strategy, adaptive reset based on similarity metrics, and deterministic degradation operators representing inverse-problem physics—may extend to other inverse problems. They specifically point to future exploration of dual-domain generalized diffusion in image and sinogram space and to possible applicability in modalities such as MRI and PET (Chen et al., 14 Aug 2025).