DM4CT: Benchmarking Diffusion Models for Computed Tomography Reconstruction
Researchers benchmark 10 diffusion models against 7 traditional methods using real-world high-energy synchrotron data.
A research team from Leiden University and TU Delft has launched DM4CT, the first comprehensive benchmark specifically designed to evaluate diffusion models for computed tomography reconstruction. Published as an ICLR 2026 submission, this systematic study addresses a critical gap in medical imaging AI by testing whether diffusion models—which have revolutionized natural image generation—can handle the practical complexities of CT reconstruction.
The benchmark includes datasets from both medical and industrial domains with challenging sparse-view and noisy configurations. Most significantly, the researchers acquired a high-resolution CT dataset from a high-energy synchrotron facility to evaluate all methods under real experimental conditions. They benchmarked ten recent diffusion-based methods against seven strong baselines, including model-based, unsupervised, and supervised approaches. The study reveals that while diffusion models show promise, they face unique challenges in CT applications including correlated noise, artifact structures, reliance on system geometry, and misaligned value ranges that don't exist in natural image domains.
All code and the real-world dataset are publicly available, providing researchers and practitioners with standardized tools to compare reconstruction methods. The analysis offers detailed insights into the behavior, strengths, and limitations of diffusion models specifically for inverse problems in medical imaging, moving beyond theoretical performance to practical deployment considerations.
- Benchmarks 10 diffusion models against 7 traditional CT reconstruction methods using real synchrotron data
- Includes medical and industrial datasets with sparse-view and noisy configurations for realistic testing
- Publicly releases codebase and high-resolution dataset to standardize evaluation in medical imaging AI
Why It Matters
Provides standardized evaluation for AI in medical imaging, helping researchers identify which models work best for real-world CT reconstruction challenges.