UniSino: CT Sinogram Standardization

• UniSino is a physics-driven foundational model that standardizes degraded CT sinograms in the projection domain to enhance reconstruction quality.
• It employs a dual-module architecture—SinoVAE for latent compression and Latent Refinement Diffusion for artifact suppression—with physics-based losses.
• Extensive evaluations across multiple datasets demonstrate its superior performance in PSNR, SSIM, and robust handling of mixed degradation scenarios.

Searching arXiv for the specified paper and closely related context. UniSino is a physics-driven foundational model for universal CT sinogram standardization that operates directly in the CT projection domain rather than the reconstructed image domain (Ai et al., 25 Aug 2025). It is designed to transform a degraded sinogram $x_{\mathrm{corr}}(s,\theta)$ into a clean, physically consistent sinogram $x_{\mathrm{std}}(s,\theta)$ that matches the distribution and physics of fully sampled, noise-free projections, with the goal of mitigating heterogeneous degradations such as undersampling and noise before they are amplified by reconstruction (Ai et al., 25 Aug 2025). In the formulation reported for UniSino, this projection-domain strategy is motivated by the observation that sinograms exhibit more uniform distributions across defect types than reconstructed images, and that correction at the raw-data stage improves downstream reconstruction quality in both single and mixed undersampling scenarios (Ai et al., 25 Aug 2025).

1. Concept and problem formulation

UniSino addresses degradation in CT raw data arising from non-standard scanning protocols and non-ideal reconstruction conditions (Ai et al., 25 Aug 2025). The degradation modes listed for the framework include sparse-view, limited-angle, low-dose, detector downsampling, ring artifacts from detector channels, geometric miscalibration, truncation, metal, and motion (Ai et al., 25 Aug 2025). In the problem setting described for the model, these degradations act on the sinogram and subsequently propagate through the nonlinear reconstruction pipeline, producing severe artifacts such as streaks, rings, structural collapse, and high-frequency distortions in reconstructed images (Ai et al., 25 Aug 2025).

The term “sinogram standardization” is used in a specific sense: transforming degraded raw projection data into a physically consistent form aligned with fully sampled, noise-free projections (Ai et al., 25 Aug 2025). This distinguishes UniSino from conventional correction procedures, which are described as artifact-specific, calibration-dependent, and limited in portability, and from image-domain foundational models, which must disentangle artifact-specific clusters in a highly nonuniform image-space distribution (Ai et al., 25 Aug 2025). The reported rationale is that a universal projection-domain model can generalize across heterogeneous degradations, including mixed artifacts, because the source-space statistics are more amenable to unified modeling (Ai et al., 25 Aug 2025).

A common misconception is to treat UniSino as merely a denoising or sparse-view completion model. The description instead positions it as a universal standardization framework spanning multiple subtasks and mixed degradation settings, with training explicitly structured to handle heterogeneous artifact types rather than a single isolated corruption mode (Ai et al., 25 Aug 2025).

2. Physical basis and mathematical structure

UniSino is grounded in standard CT forward physics, beginning with the Beer–Lambert law for x-ray transmission (Ai et al., 25 Aug 2025):

$$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$

where $\mu(l)$ is the linear attenuation coefficient along path $\mathcal{L}$, $I$ is the detected intensity, and $I_0$ is the incident intensity (Ai et al., 25 Aug 2025). After logarithmic transformation, the projection value becomes a line integral,

$p = -\ln(I/I_0) = \int_{\mathcal{L}} \mu(l)\,dl,$

which is the basis of the sinogram representation used by the model (Ai et al., 25 Aug 2025).

The framework further adopts the Radon-transform view of projection data (Ai et al., 25 Aug 2025):

$$Rf(s,\theta) = \int_{\mathbb{R}^2} f(\mathbf{x})\,\delta\!\big(s - \mathbf{x}\cdot\mathbf{n}_\theta\big)\,d\mathbf{x},$$

with unit normal $\mathbf{n}_\theta = (\cos\theta, \sin\theta)$ (Ai et al., 25 Aug 2025). Under low-dose acquisition, photon counts are modeled as Poisson random variables. If $x_{\mathrm{std}}(s,\theta)$0 are expected counts, then the measured counts satisfy

$x_{\mathrm{std}}(s,\theta)$1

and the log-domain measurement is

$x_{\mathrm{std}}(s,\theta)$2

yielding heteroscedastic noise in the projection domain (Ai et al., 25 Aug 2025). In the reported training pipeline, low-dose artifacts are simulated precisely by sampling $x_{\mathrm{std}}(s,\theta)$3 from $x_{\mathrm{std}}(s,\theta)$4 and then applying the log transform (Ai et al., 25 Aug 2025).

Rather than imposing the full Helgason–Ludwig moment conditions, UniSino uses two practical projection-domain consistency constraints through SinoLoss: cross-view mean consistency and angular visibility constraints (Ai et al., 25 Aug 2025). Their concrete forms are given as

$x_{\mathrm{std}}(s,\theta)$5

and

$x_{\mathrm{std}}(s,\theta)$6

where $x_{\mathrm{std}}(s,\theta)$7 is the cross-angle mean derived from observed angles and $x_{\mathrm{std}}(s,\theta)$8 is a zero–one visibility mask inferred through intersection, backprojection, and forward projection (Ai et al., 25 Aug 2025). SinoLoss is described as combining these constraints with bounded-variation behavior to discourage nonphysical oscillations (Ai et al., 25 Aug 2025).

This suggests that UniSino’s “physics-driven” characterization is not limited to using a CT forward model in data simulation; it also embeds physically motivated regularity conditions directly into the training objective.

3. Architecture: SinoVAE and latent refinement diffusion

UniSino comprises two tightly coupled modules that both operate in the projection domain: SinoVAE and Latent Refinement Diffusion (LRD) (Ai et al., 25 Aug 2025). SinoVAE is described as a Sinogram Variation Autoencoder with perceptual compression, dual pathways, and physics-driven constraints (Ai et al., 25 Aug 2025). Its encoder $x_{\mathrm{std}}(s,\theta)$9 produces the mean and variance of a latent Gaussian,

$$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$0

with KL regularization used to structure the latent space (Ai et al., 25 Aug 2025).

The decoder stage is explicitly dual-path (Ai et al., 25 Aug 2025). Decoder 1 reconstructs full-frequency content, while Decoder N emphasizes artifact-sensitive high-frequency structure such as streaks, rings, and undersampling edges, with high-frequency guidance computed using Sobel gradients:

$$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$1

Discriminator-based adversarial heads are attached to both decoders to sharpen realism and suppress artifacts (Ai et al., 25 Aug 2025). The latent embedding is therefore decomposed into a full-frequency pathway preserving global structure and intensity, and a high-frequency pathway encoding artifact-sensitive details (Ai et al., 25 Aug 2025).

The second module, LRD, is a conditional diffusion model operating in latent space to refine SinoVAE latents of undersampled sinograms (Ai et al., 25 Aug 2025). Its forward process over $$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$2 steps is

$$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$3

with $$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$4 and $$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$5 (Ai et al., 25 Aug 2025). The reverse denoising process conditioned on the corrupted-sinogram latent is

$$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$6

where $$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$7 predicts the added noise and $$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$8 is the conditioning latent (Ai et al., 25 Aug 2025). Conditioning uses both full-frequency and high-frequency latent channels so that restoration remains faithful to the corrupted input while suppressing artifacts (Ai et al., 25 Aug 2025).

A plausible implication is that the division of labor between SinoVAE and LRD is central to the model’s reported efficiency: compression and feature structuring occur before diffusion, rather than performing unconditional generation directly in full-resolution sinogram space.

4. Training procedure and artifact simulation

UniSino is trained in two stages using self-supervised data generation from clean sinograms (Ai et al., 25 Aug 2025). In Stage I, SinoVAE is trained on standardized or clean sinograms with simulated degraded counterparts (Ai et al., 25 Aug 2025). The objective is

$$I = I_0 \exp\!\left(-\int_{\mathcal{L}} \mu(l)\,dl\right),$$9

Here $\mu(l)$0 denotes the ground-truth sinogram and $\mu(l)$1 the decoder (Ai et al., 25 Aug 2025). In Stage II, LRD is trained in latent space with conditioning from the undersampled-sinogram latent $\mu(l)$2 using

$\mu(l)$3

After $\mu(l)$4 diffusion steps, the refined latent $\mu(l)$5 is decoded by the SinoVAE global decoder (Ai et al., 25 Aug 2025).

The degradation simulation protocol covers multiple artifact modes (Ai et al., 25 Aug 2025). Sparse-view corruption is modeled by removing angle subsets, limited-angle by restricting the angular range, truncation by cropping the detector axis and setting values outside the field of view to nominal values, downsampling by sub-sampling detector channels, ring artifacts by adding channel-specific offsets, geometry artifacts by angle-dependent detector shifts, metal artifacts by modifying projected high-attenuation regions, and motion artifacts by forward projecting a deformed object (Ai et al., 25 Aug 2025). Mixed degradation is created by random degradation mixing during training, which is explicitly reported as a mechanism for robustness to co-occurring artifacts (Ai et al., 25 Aug 2025).

Training was conducted primarily on the NLST lung dataset with 300 patients and 800,112 slices, of which 790,112 were used for training and 10,000 for testing; sinogram size was $\mu(l)$6 (Ai et al., 25 Aug 2025). The implementation used PyTorch with Adam optimizer on two NVIDIA RTX 4090 GPUs, with learning rates of $\mu(l)$7 for SinoVAE and $\mu(l)$8 for LRD (Ai et al., 25 Aug 2025). SinoVAE and LRD were trained sequentially, with the encoder frozen during diffusion training (Ai et al., 25 Aug 2025).

5. Datasets, tasks, and empirical performance

Although UniSino is described as being validated across eight CT datasets, the core benchmarking is reported on four representative datasets: NLST, CQ500, LIDC-IDRI, and KiTS19 (Ai et al., 25 Aug 2025). Additional generalization tests use CHAOS, QIN LUNG, LiTS, and MSD Colon (Ai et al., 25 Aug 2025). The subtasks include sparse-view completion, limited-angle restoration, low-dose denoising, downsample restoration, ring artifact correction, truncation correction, geometry correction, and mixed-artifact restoration (Ai et al., 25 Aug 2025). Evaluation uses projection-domain PSNR, SSIM, and NRMSE on standardized sinograms, with image-domain metrics obtained after applying FBP to the standardized sinograms (Ai et al., 25 Aug 2025).

The overall projection-domain comparison reported in Table I gives UniSino PSNR 45.522, SSIM 96.272, and NRMSE 0.00607 (Ai et al., 25 Aug 2025). The corresponding values reported for CycleGAN are 27.398, 77.046, and 0.08221; for ViT, 27.881, 70.363, and 0.06632; for U-Net, 39.764, 91.087, and 0.01820; and for DDPM, 39.613, 95.432, and 0.01385 (Ai et al., 25 Aug 2025).

Per-task projection-domain results are also reported (Ai et al., 25 Aug 2025):

Subtask	PSNR	SSIM
SV	48.938	97.331
LA	39.794	93.237
LD	49.420	97.305
DS	47.567	95.091
RI	49.266	97.290
TR	42.867	96.391
GE	40.800	97.262

The corresponding NRMSE values are reported as 0.00353 for SV, 0.01141 for LA, 0.00335 for LD, 0.00373 for DS, 0.00341 for RI, 0.00794 for TR, and 0.00912 for GE (Ai et al., 25 Aug 2025). Cross-dataset generalization on NLST evaluation is reported as PSNR 46.19, SSIM 96.731, and NRMSE 0.00331, exceeding the reported results for CycleGAN, ViT, U-Net, and DDPM (Ai et al., 25 Aug 2025).

For mixed undersampling examples combining LD+RI+SV and TR+GE+LA, the reported PSNR values are 42.641 and 37.152, respectively (Ai et al., 25 Aug 2025). The text attributes these outcomes to random mixing during training and projection-domain consistency, which together yield reliable normalization under compounded artifacts (Ai et al., 25 Aug 2025).

6. Ablation, deployment characteristics, and limitations

Ablation studies identify the physics-driven components and projection-domain design as critical to performance (Ai et al., 25 Aug 2025). Replacing SinoVAE with a standard encoder–decoder reduces projection-domain performance to PSNR 43.647, SSIM 94.483, and NRMSE 0.00791 (Ai et al., 25 Aug 2025). Removing SinoLoss yields PSNR 44., SSIM 95.612, and NRMSE 0.00654 (Ai et al., 25 Aug 2025). Full UniSino is reported at PSNR 45.522, SSIM 96.272, and NRMSE 0.00607 (Ai et al., 25 Aug 2025). The associated interpretation in the source is that explicit high-frequency preservation and physics-guided losses reduce artifact propagation after backprojection and enhance restoration fidelity (Ai et al., 25 Aug 2025).

The model is described as efficient at inference because SinoVAE compresses sinograms into compact latents and the diffusion stage operates conditionally in latent space rather than as unconditional full-resolution generation (Ai et al., 25 Aug 2025). Exact throughput numbers are not reported, but the model is described as having substantially shorter reconstruction times than patch-based baselines and unconditional DDPM (Ai et al., 25 Aug 2025). Standardized sinograms can be fed into standard FBP or iterative reconstruction, and no specialized post-processing is required (Ai et al., 25 Aug 2025).

The paper also introduces the SimSinoCT dataset for benchmarking sinogram standardization and releases code at the stated repository (Ai et al., 25 Aug 2025). This reinforces the model’s role not only as a restoration architecture but as part of a broader evaluation framework for projection-domain standardization.

Reported failure modes include extreme 2D truncation outside inferred visibility masks and very large miscalibrations where simple geometric shifts are insufficient (Ai et al., 25 Aug 2025). The text suggests potential benefits from explicit scanner geometry modeling in future work (Ai et al., 25 Aug 2025). Other future directions listed are extending to 3D sinograms for volumetric CT, cross-modality adaptation to MRI k-space and PET sinograms, and multi-domain strategies combining standardized sinograms with image-domain constraints (Ai et al., 25 Aug 2025).

7. Position within CT reconstruction research

UniSino is positioned as a projection-domain foundational model, in contrast to foundational models that operate in image space (Ai et al., 25 Aug 2025). The distinction is methodological as well as conceptual. Image-domain systems must address artifacts after reconstruction, whereas UniSino attempts to standardize the raw measurement representation before backprojection or iterative inversion (Ai et al., 25 Aug 2025). In the reported framing, this avoids amplification of source-space inconsistencies and exploits the more uniform distribution of sinograms across degradation types (Ai et al., 25 Aug 2025).

Its broader significance lies in harmonizing raw CT data across protocols and scanners while reducing dependence on vendor-specific, artifact-specific preprocessing (Ai et al., 25 Aug 2025). This suggests a shift in medical imaging foundation-model design from image-domain semantics toward raw-data modeling with embedded acquisition physics. A plausible implication is that UniSino belongs to a class of models that treat the measurement domain itself as the principal locus of generalization, rather than as a preliminary step to be discarded once reconstruction begins.

Within that framing, UniSino can be understood as a universal standardization system for CT sinograms whose core contributions are a physics-constrained projection-domain representation, a dual-path variational latent model, and latent diffusion refinement trained under mixed degradation regimes (Ai et al., 25 Aug 2025). Its reported empirical performance, cross-dataset generalization, and ablation results collectively support the claim that projection-domain foundational modeling is a viable strategy for robust CT raw-data enhancement (Ai et al., 25 Aug 2025).