PaperlayerView exact revision on arXiv

Source-linked AI summary

DM4CT: Benchmarking Diffusion Models for Computed Tomography Reconstruction

Jiayang Shi, Daniel M. Pelt, K. Joost Batenburg

arXiv:2602.18589v1eess.IVcs.AIcs.CV

TL;DR

Practical CT reconstruction differs from the idealized linear inverse-problem setting because noise, artifacts, sparse views, geometry, and value-range mismatches complicate diffusion-prior deployment. DM4CT addresses this gap with a systematic benchmark spanning medical, industrial, and synchrotron datasets and comparing diffusion methods with established baselines. The results show that diffusion models can be strong, competitive priors, but no single method consistently dominates and real-world performance is generally worse than simulated performance.

  • Problem

    Practical CT pipelines depart from idealized linear models because of complex noise, nonlinear preprocessing, and artifacts, motivating comprehensive evaluation of diffusion reconstruction methods.

  • Method

    DM4CT benchmarks diffusion-based CT reconstruction methods against classical, model-based, unsupervised, and supervised approaches across controlled simulated datasets and a real-world synchrotron dataset.

  • Results

    Diffusion models generally outperform classical and model-based reconstruction methods, show advantages under sparse-view and high-noise conditions, but no single diffusion method consistently leads across datasets and configurations.

  • Takeaways & Limitations

    DM4CT provides practical evidence that diffusion models can serve as strong CT reconstruction priors while exposing challenges in balancing learned priors with measurement consistency under realistic conditions.

  • Takeaways & Limitations

    Real-world performance is generally worse than simulated-data performance, and practical deployment remains constrained by value-range misalignment and computational overhead from acquisition geometry.

Abstract

from arXiv · show

Diffusion models have recently emerged as powerful priors for solving inverse problems. While computed tomography (CT) is theoretically a linear inverse problem, it poses many practical challenges. These include correlated noise, artifact structures, reliance on system geometry, and misaligned value ranges, which make the direct application of diffusion models more difficult than in domains like natural image generation. To systematically evaluate how diffusion models perform in this context and compare them with established reconstruction methods, we introduce DM4CT, a comprehensive benchmark for CT reconstruction. DM4CT includes datasets from both medical and industrial domains with sparse-view and noisy configurations. To explore the challenges of deploying diffusion models in practice, we additionally acquire a high-resolution CT dataset at a high-energy synchrotron facility and evaluate all methods under real experimental conditions. We benchmark ten recent diffusion-based methods alongside seven strong baselines, including model-based, unsupervised, and supervised approaches. Our analysis provides detailed insights into the behavior, strengths, and limitations of diffusion models for CT reconstruction. The real-world dataset is publicly available at zenodo.org/records/15420527, and the codebase is open-sourced at github.com/DM4CT/DM4CT.

1 INTRODUCTION

DM4CT addresses the practical difficulty of applying diffusion priors to CT by providing a systematic benchmark across simulated and real-world settings. It compares diffusion methods with established approaches while organizing them by data-consistency and prior-integration strategies.

  • Motivation: CT reconstruction is ill-posed under sparse or noisy measurements, so methods incorporate priors ranging from total variation to supervised learned models.Multiple reconstructions can fit indirect measurements equally well, motivating prior knowledge.
  • Motivation: Practical CT pipelines depart from idealized linear models because of complex noise, nonlinear log transformations, and artifacts.These factors motivate realistic evaluation rather than direct transfer from natural-image diffusion settings.
  • Benchmark: DM4CT benchmarks diffusion-based CT reconstruction methods against strong established baselines using medical, industrial, and synchrotron-acquired datasets.The benchmark includes controlled noise and artifact conditions and a real-world dataset acquired from rock samples.
  • Contributions: DM4CT contributes a unified taxonomy of diffusion approaches based on how they incorporate data consistency and prior knowledge.The authors also implement benchmarked methods in the diffusers framework and release the codebase.
  • Benchmark: The benchmark evaluates robustness to limited views, noise, and ring artifacts using qualitative visual assessment and quantitative image-quality and computational-efficiency metrics.Its overview applies representative methods to measured sinograms under a shared forward model.
  • Contributions: The authors position DM4CT as a systematic benchmark rather than a new reconstruction algorithm.Its purpose is to provide comparative evidence about diffusion models in CT.

2 PRELIMINARIES

CT reconstruction recovers an unknown object from indirect projection measurements, while diffusion models provide pixel- or latent-space priors for generating plausible images. Their use in inverse problems requires conditioning the reverse denoising process on measurements and enforcing suitable priors.

  • Computed Tomography: CT models projection measurements as y = Ax, where A is determined by acquisition geometry and sparse or noisy measurements make reconstruction underdetermined and ill-posed.The observed measurements differ from ideal measurements by measurement noise.
  • Computed Tomography: Classical total-variation priors assume image smoothness, whereas supervised and implicit neural priors provide more expressive or task-specific alternatives.The alternatives include paired-image training, Deep Image Prior, and implicit neural representations.
  • Diffusion Models: Pixel-space diffusion gradually adds Gaussian noise and trains a neural network to approximate the score function for reverse denoising.The forward and backward processes are described using stochastic differential equations.
  • Diffusion Models: Latent diffusion performs forward diffusion and reverse denoising in a lower-dimensional latent space before decoding the result back into data space.The benchmark uses a VQ-VAE as the encoder-decoder pair.

3 DM4CT

DM4CT organizes diffusion-based CT reconstruction methods around how they incorporate measurement consistency and prior knowledge, then evaluates them across medical, industrial, and synchrotron datasets. The benchmark uses shared pixel- and latent-space diffusion backbones and compares methods under varied noise, sparsity, artifacts, and real-world conditions.

  • Diffusion models for CT reconstruction: Conditional diffusion reconstruction combines the learned score with a measurement-informed correction, approximated through clean-image estimates during reverse denoising.The conditional term is generally intractable because measurements depend on the clean image rather than the noised state.
  • Reconstruction strategies: DM4CT classifies methods by data-consistency gradient steering, optimization steps, plug-and-play priors, variational Bayesian inference, and approximate pseudoinverse guidance.Approximate pseudoinverses use FBP or SIRT because directly computing the Moore–Penrose inverse is generally infeasible for CT systems.
  • Datasets and configurations: The benchmark evaluates medical, industrial, and high-resolution synchrotron CT, including controlled sparse-view, noise, artifact, and real experimental settings.The synchrotron data comprise two rock samples and support slice-wise 2D reconstruction because of their parallel-beam, circular-trajectory geometry.
  • Implementation and comparison methods: Each dataset uses one pixel-space and one latent-space diffusion model as shared backbones for all diffusion methods, supporting fair method comparisons.Table 2 reports PSNR and SSIM across medical, industrial, and synchrotron configurations; methods exceeding 40 GB GPU memory are not executed.

4 RESULTS AND DISCUSSIONS

DM4CT finds that diffusion methods often outperform classical and MBIR approaches, especially under sparse or noisy measurements, but no diffusion subclass dominates across all settings. Results also expose trade-offs involving perceptual fidelity, data consistency, computational resources, and real-world generalization.

  • Reconstruction performance: Diffusion methods generally outperform classical and MBIR approaches in PSNR and SSIM but often trail fully supervised SwinIR.INR can achieve comparable metrics, particularly without noise and on the real-world dataset.
  • Cross-setting robustness: No diffusion method or subclass consistently outperforms the others across datasets and configurations, and real-world performance is generally worse than simulated-data performance.The paper suggests limited training-data quality and distribution shift as likely contributors to the real-world degradation.
  • Prior and data consistency: Increasing DPS’s data-consistency step size initially improves data fit and reconstruction quality, but overly large values disrupt denoising and cause collapse into measurement-noise-dominated reconstructions.This illustrates the need to balance prior knowledge against measurement fidelity.
  • Latent diffusion: Explicit optimization steps such as ReSample correct latent-space discontinuities in noise-free settings but can overfit noisy measurements and amplify noise-like features.Gradient-only enforcement in latent diffusion is associated with discontinuities even under noise-free conditions.
  • Sparsity and noise: Pixel-space diffusion shows clearer advantages as projection views become sparse or measurement noise increases, while the gap narrows with cleaner or denser measurements.The benchmark attributes this pattern to the strength of learned priors under limited or corrupted measurement information.
  • Computational efficiency: Pixel diffusion is generally more memory- and time-efficient at inference than latent diffusion, but training costs and memory requirements vary substantially across methods.SwinIR is fastest at inference but memory-intensive; INR and DIP use less memory but are slower, while latent diffusion adds VQ-VAE training.

5 CONCLUSION

DM4CT finds that diffusion models can be strong priors for CT reconstruction, but practical deployment remains challenged by balancing learned priors with measurement consistency under realistic conditions. The benchmark is intended to support future research and narrow the gap between methodological development and practical applicability.

  • Diffusion models can serve as strong priors and achieve competitive performance across varied CT reconstruction scenarios.
  • Balancing learned priors with measurement data consistency remains difficult under noise, artifacts, and sparse-view measurements.
  • DM4CT is positioned as a resource for advancing diffusion-based inverse-problem research and practical applicability.
  • The study releases datasets, code, and documentation to support reproducibility.

A.1 LIMITATIONS

The benchmark has limitations involving forward-operator accuracy, hyperparameter selection, domain coverage, evaluation metrics, and the preliminary nature of downstream segmentation analysis.

  • The benchmark assumes accurate and known forward operators, excluding mechanical misalignments and calibration errors that may affect real deployments.
  • Method-specific hyperparameters may remain suboptimal across methods and scenarios despite grid search and additional optimization.
  • Medical and industrial datasets cover only a subset of real-world CT applications, limiting generalization to other domains and imaging tasks.
  • Reliance primarily on PSNR and SSIM may miss practical reconstruction fidelity, particularly when image intensity ranges vary.
  • The exploratory segmentation study is preliminary because results depend on the SAM model and instance-level mask-matching strategy.

A.2 BROADER IMPACT

Diffusion reconstruction can fill information gaps using prior knowledge, but this may introduce content not grounded in measurements. The paper therefore emphasizes rigorous clinical and domain-specific validation before practical deployment.

  • Diffusion priors can fill information gaps but may introduce content not grounded in measurement data, creating a hallucination concern.
  • Medical deployment requires rigorous clinical validation, while industrial use requires careful verification and domain-specific assessment for reliability and safety.

A.3 USE OF LARGE LANGUAGE MODELS

The authors used ChatGPT to refine text at the sentence level, including grammar correction and rephrasing.

  • ChatGPT assisted with sentence-level grammar correction and rephrasing.

A.4 RANGE NULL SPACE DECOMPOSITION

The range-null decomposition separates reconstruction content supported by measurements from content introduced by priors. Because explicit pseudoinverse computation is impractical for CT, the benchmark estimates the null component iteratively with Landweber updates.

  • Decomposition: The range component A†Ax is data-consistent, whereas the null component (I − A†A)x contains information not constrained by measurements.Different null components can yield multiple feasible reconstructions for the same measurement.
  • Purpose: The decomposition provides a structured way to evaluate data consistency and prior influence in diffusion-based CT reconstruction.
  • Practical computation: Directly forming the pseudoinverse A† is generally infeasible because CT system matrices are high-dimensional and sparse.
  • Practical computation: Landweber iteration estimates the null-space component without explicitly constructing A†, using the forward operator, object, step size, and tolerance.
  • Practical computation: The iterative procedure initializes x and r = Ax, then updates x using αA^T r until the residual norm falls below ε.

A.5 DIFFUSION MODELS FOR CT RECONSTRUCTION

Diffusion-based CT reconstruction commonly embeds data consistency within reverse diffusion. Each timestep estimates a clean image, applies a consistency refinement, and continues the denoising trajectory.

  • Reverse diffusion with data consistency: At each reverse-diffusion timestep, the method estimates a clean image from the noisy sample and applies a data-consistency gradient.This guides the sample trajectory toward agreement with measured data.
  • Reverse diffusion with data consistency: The resulting iterative template combines denoising updates with measurement-based refinement throughout reconstruction.

A.6 EXPERIMENTAL SETUPS FOR MEDICAL AND INDUSTRIAL DATASETS

The benchmark combines controlled medical and industrial CT experiments with simulated noise and artifacts. Its diffusion template starts from Gaussian noise, denoises using a pretrained score estimator, and applies data consistency at each step.

  • Dataset splits: The medical dataset uses volumes L067, L096, L109, L143, L192, L286, L291, L310, and L333 for training, reserving L506 for testing.
  • Diffusion reconstruction: The diffusion reconstruction template initializes x_T from N(0, I), estimates noise and a clean image, performs a reverse scheduler step, and applies data consistency.
  • Dataset splits: The LoDoInd dataset uses slices 501–3000 for training and 3001–3500 for testing from the selected 4,000-slice set.
  • Normalization: Medical and industrial volumes are linearly mapped to [−1, 1], using HU bounds for medical CT and full-volume extrema for industrial CT.
  • Noise and artifacts: Poisson projection noise is simulated through the Beer–Lambert relationship, random photon counts, and inversion back to noisy projections.
  • Noise and artifacts: Ring artifacts are simulated by adding fixed-pattern noise to randomly selected detector columns, producing column-wise sinogram inconsistencies.
  • Noise and artifacts: The configurations vary projection-angle counts, noise severity, and ring-artifact severity to test robustness under increasing data corruption.

A.7 REAL-WORLD SYNCHROTRON CT DATASET

The real-world synchrotron setting exposes value-range mismatch, limited training data, and distribution shift as practical challenges for diffusion-based CT reconstruction. Results show dataset- and method-dependent trade-offs between metrics, detail, geometry, and robustness.

  • Practical deployment challenges: Industrial value ranges depend strongly on scanning conditions, while medical raw correction factors may be inaccessible, creating diffusion-model input inconsistencies.
  • Practical deployment challenges: Linear range alignment yields approximate reconstructions from 40-angle raw projections while recovering the main structural features.
  • Noise-model mismatch: Under matched FBP noise levels, DDS reconstructs fine structures under Gaussian noise but degrades substantially under Poisson noise.
  • Downstream evaluation: The segmentation analysis is preliminary rather than clinical, because outcomes depend on SAM and mask matching and require more controlled studies.
  • Autoencoder evaluation: VQ-VAE provides the strongest representation and CT reconstruction quality among the evaluated autoencoder variants, while natural-image autoencoder adaptation remains preliminary.
  • Real-world results: On the real-world dataset, diffusion methods generally perform worse than on simulated data, while SwinIR often scores higher but produces smoother reconstructions.
  • Real-world results: DPS and PGDM produce plausible contours with slight distortions, MCG introduces unnatural porosities, PSLD misses geometry, and Reddiff struggles with fine textures.
Loading 2602.18589v1…