⚡ electrical engineering

Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction for Sparse-View CT

This paper proposes the Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction (CDPIR) framework, which integrates a Scalable Interpolant Transformer (SiT) with model-based iterative reconstruction to leverage cross-distribution priors and significantly improve sparse-view CT reconstruction quality, particularly in out-of-distribution scenarios.

Original authors: Haodong Li, Shuo Han, Haiyang Mao, Yu Shi, Changsheng Fang, Jianjia Zhang, Weiwen Wu, Hengyong Yu

Published 2026-04-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Haodong Li, Shuo Han, Haiyang Mao, Yu Shi, Changsheng Fang, Jianjia Zhang, Weiwen Wu, Hengyong Yu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a giant, complex jigsaw puzzle, but someone has stolen 90% of the pieces. You only have a few scattered pieces left, and you need to figure out what the whole picture looks like.

This is exactly the challenge doctors face with Sparse-View CT scans. To save time and reduce radiation exposure, they take fewer "pictures" (projections) of the patient's body. The problem? When you try to reconstruct the full image from so few pieces, you get a blurry, streaky mess full of artifacts. It's like trying to guess the face of a stranger based on just their left ear and a single tooth.

For years, computers have tried to "hallucinate" (guess) the missing parts using AI. But here's the catch: If the AI was trained on pictures of adults, and you ask it to reconstruct a picture of a baby, it might get confused. It might try to force adult features onto the baby, or get stuck because the "rules" of the puzzle changed. This is called the Out-of-Distribution (OOD) problem.

Enter the authors of this paper, who propose a new system called CDPIR. Here is how it works, using some everyday analogies:

1. The "Universal Translator" (Cross-Distribution Priors)

Most AI models are like specialists who only speak one dialect. If you speak to them in a different accent, they misunderstand.

The Old Way: Train a model on "Siemens scanners" and "Abdomen scans." If you give it a "GE scanner" and a "Heart scan," it fails.
The CDPIR Way: The authors built a "Universal Translator." They trained their AI on a mix of many different scanners, body parts, and protocols all at once.
The Magic Trick (Classifier-Free Guidance): Imagine the AI has two modes:
1. The "Anatomy" Mode: It learns the universal rules of the human body (e.g., "The heart is always in the chest," "Bones are hard"). This part is learned by occasionally turning off the "instruction manual" (the specific scanner type) during training.
2. The "Style" Mode: It learns the specific look of different scanners (e.g., "This scanner makes bones look grainy," "That scanner makes soft tissue look smooth").
  By combining these, the AI knows the structure of the body regardless of the scanner, but can adjust the texture to match the specific machine taking the picture.

2. The "Smart Referee" (Iterative Reconstruction)

Imagine the AI is a painter trying to fix a damaged painting.

The Diffusion Part: The AI starts with a canvas full of static noise (like TV snow) and slowly "denoises" it, turning the snow into a clear image. It's like sculpting a statue out of a block of marble, chipping away the noise to reveal the shape.
The Problem: Sometimes the AI gets too creative and invents things that aren't there (hallucinations).
The Referee (Data Consistency): This is where the "Iterative" part comes in. After the AI makes a guess, a strict "Referee" (based on physics) checks: "Does this guess actually match the few X-ray pieces we have?"
- If the AI guessed a bone where there shouldn't be one, the Referee says, "No, that contradicts the X-ray data," and pushes the image back toward the truth.
- The AI and the Referee take turns: The AI adds detail, the Referee corrects the physics. They keep swapping until the image is both detailed and physically accurate.

3. The "High-Speed Train" (Scalable Interpolant Transformer)

Previous AI models were like old, clunky trains that took a long time to get from "Noise" to "Clear Image." They often got stuck in the middle or took a winding, inefficient path.

The New Engine: The authors used a new architecture called SiT (Scalable Interpolant Transformer). Think of this as a high-speed bullet train. It calculates the most direct, smooth path from the noise to the final image.
Why it matters: It gets to the solution faster, with fewer stops, and it's less likely to crash (produce errors) when the terrain changes (different scanners).

The Result: A Super-Resilient Doctor's Assistant

When the researchers tested CDPIR:

It didn't panic when moved from one type of hospital scanner to another.
It didn't hallucinate fake tumors or bones.
It preserved fine details (like tiny blood vessels) that other methods smoothed over.
It worked on real patients, not just computer simulations, even with extremely low radiation doses.

In a nutshell: CDPIR is a smart, adaptable AI that learns the "universal grammar" of human anatomy while respecting the "dialect" of the specific machine taking the photo. It then works in a team with physics-based rules to ensure the final picture is both beautiful and medically accurate, even when the input data is incomplete or comes from an unfamiliar source. It's a major step toward making safe, low-radiation CT scans a standard part of healthcare.

1. Problem Statement

Sparse-View Computed Tomography (SVCT) is a critical technique for reducing radiation dose and improving temporal resolution in clinical settings (e.g., cardiac and extremity imaging). However, reconstructing images from limited projections is an ill-posed inverse problem that leads to severe streak artifacts and loss of fine details.

The paper identifies two primary challenges hindering the clinical deployment of current deep learning-based SVCT methods:

Sparse-View Artifacts: Undersampled projections cause structural fidelity loss and noise.
Out-of-Distribution (OOD) Shifts: Real-world deployment involves significant distribution shifts between training and testing data due to different scanners (e.g., Siemens vs. GE), protocols, anatomical foci (e.g., abdomen vs. heart), and noise characteristics.
- Existing discriminative models (CNNs, Transformers) and standard diffusion models often fail under OOD conditions, leading to hallucinations (visually plausible but anatomically incorrect structures) or severe performance degradation.
- Current solutions like Test-Time Adaptation (TTA) incur high computational overhead, while large-scale models struggle to control epistemic uncertainty in rare OOD cases.

2. Methodology: CDPIR Framework

The authors propose Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction (CDPIR), a framework that unifies a scalable generative prior with physics-based iterative reconstruction.

A. Core Architecture: Scalable Interpolant Transformer (SiT)

Instead of standard Diffusion Transformers (DiT) that learn score functions, CDPIR utilizes SiT, which learns a velocity field under a stochastic interpolant formulation.

Stochastic Interpolant: Defines a linear interpolation between clean data ( $x_0$ ) and noise ( $\epsilon$ ): $x_t = \alpha_t x_0 + \sigma_t \epsilon$ .
Advantage: This formulation provides more explicit control over the transition from noise to data, supporting both stochastic (SDE) and deterministic (ODE) sampling with improved numerical stability and convergence efficiency compared to Variance Preserving (VP) SDEs.

B. Cross-Distribution Prior via Classifier-Free Guidance (CFG)

To address OOD shifts, CDPIR employs a Classifier-Free Guidance (CFG) strategy during training:

Training Mechanism: The model is trained on multiple datasets (e.g., AAPM, COCA) using domain tokens. Crucially, the domain token is randomly dropped (replaced with a null token $\emptyset$ ) with a fixed probability.
Feature Disentanglement: This forces the model to learn two complementary representations:
1. Domain-Invariant Prior ( $v_{inv}$ ): Learned from the null-conditioned branch, capturing shared anatomical structures (e.g., heart, lungs, spine) across all domains.
2. Domain-Specific Residual ( $v_{spec}$ ): Learned from the conditional branch, capturing appearance variations (texture, noise, edge sharpness) specific to a scanner or protocol.
Inference: During sampling, the model combines these priors. The unconditional branch ensures robust anatomical structure, while the conditional branch (if a matching token is available) refines domain-specific details. If no token matches, the model relies on the robust unconditional prior.

C. Residual-Guided Iterative Reconstruction

CDPIR integrates the diffusion prior into a Model-Based Iterative Reconstruction (MBIR) loop to enforce data consistency:

Alternating Solver: The algorithm alternates between:
1. Diffusion Prior Step: Uses the SiT model to denoise the current estimate based on the cross-distribution prior.
2. Data Consistency Step: Uses an improved Adaptive Steepest Descent Projection Onto Convex Sets (ASD-POCS) with Total Variation (TV) regularization. This step projects the image back onto the manifold of measurements ( $Ax \approx y$ ) to suppress artifacts and prevent hallucinations.
Formulation: The problem is solved as a constrained optimization minimizing a combination of data fidelity, TV regularization, and the diffusion prior objective.

3. Key Contributions

Unified Framework: CDPIR is the first to integrate conditional diffusion, stochastic interpolant sampling, and residual-guided updates into a single iterative reconstruction pipeline for SVCT.
Cross-Distribution Generalization: By leveraging CFG with random token dropping, the method learns a transferable prior that separates anatomy (domain-invariant) from appearance (domain-specific), achieving state-of-the-art (SOTA) performance on unseen distributions without retraining.
Scalable Backbone: The use of SiT allows for systematic model scaling and better handling of long-range dependencies compared to CNN-based diffusion models.
Robustness: The framework effectively suppresses sparse-view artifacts while preventing prior-driven hallucinations in challenging OOD scenarios.

4. Experimental Results

The authors validated CDPIR on five datasets (AAPM, Stanford COCA, XCAT, GE Clinical, MARS PCCT) covering diverse scanners, protocols, and anatomical regions.

Quantitative Performance:
- CDPIR significantly outperformed SOTA baselines (DDS, DiffusionMBIR, DDIP3D, RED-CNN, Restormer) in PSNR, SSIM, and LPIPS.
- OOD Robustness: In fixed-test OOD evaluations (e.g., training on AAPM, testing on XCAT), CDPIR showed near-zero performance degradation ( $\Delta$ PSNR $\approx$ 0), whereas CNN and Transformer baselines suffered significant drops (e.g., >4 dB PSNR loss).
- Clinical Zero-Shot: On unseen clinical datasets (GE Cardiac and MARS PCCT) without fine-tuning, CDPIR achieved superior structural similarity (SSIM +0.07 over DDS) and artifact suppression.
Ablation Studies:
- Backbone: Transformer (SiT) significantly outperformed U-Net baselines in OOD settings.
- CFG: Removing the conditional-drop mechanism degraded performance, confirming the necessity of learning a shared structural prior.
- Stochastic Interpolant: SiT demonstrated smoother convergence and better compatibility with data consistency steps compared to VP-SDE formulations.
- Model Capacity: Larger models (CDPIR-B) yielded better results, but even smaller variants outperformed baselines.
Efficiency: While diffusion methods are inherently iterative, CDPIR converges faster than competitors. Using 300 steps instead of 1,000 maintained high quality while reducing inference time significantly (approx. 8.5 mins per scan vs. 23 mins for DDIP3D).

5. Significance and Impact

Clinical Relevance: CDPIR addresses the critical "domain shift" problem, making AI-based SVCT reconstruction viable for diverse clinical environments (different scanners, protocols, and body parts) without the need for per-site retraining or expensive test-time adaptation.
Safety: By explicitly coupling generative priors with physics-based data consistency, the method mitigates the risk of hallucinations, a major safety concern in medical imaging where false structures could mimic pathology.
Foundation Model Potential: The ability to train a single model across heterogeneous datasets positions CDPIR as a potential "foundation model" for medical image reconstruction, capable of generalizing to new scanners and tasks.
Future Directions: The paper suggests that this approach could be extended to other modalities and combined with implicit neural representations (INRs) or 3D Gaussian splatting for further advancements in tomographic reconstruction.

In conclusion, CDPIR represents a significant step forward in making sparse-view CT reconstruction robust, accurate, and clinically deployable across diverse and challenging real-world scenarios.