EUVIMEDEUVIMED entdecken
Crossref

Physics-Informed Generative AI for Low-Dose CT Reconstruction: From Iterative Reconstruction to Diffusion Models

Byunghoon Kang, Hwunjae Lee, Gaehwan Jin, V. R. Singh

Journal of Medical Imaging · 30.06.2026 · 9 (1) · 81-93

DOI: 10.31916/sjmi2026-01-9 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Medical Imaging, online erschienen am 30.06.2026. Verfasst von Byunghoon Kang und 3 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Purpose

Computed tomography (CT) has become indispensable across nearly all clinical specialties owing to its rapid acquisition, high spatial resolution, and broad diagnostic applicability. However, concerns regarding cumulative radiation exposure remain a major challenge, particularly in oncologic surveillance, pediatric imaging, lung cancer screening, and cardiovascular examinations. Low-dose CT (LDCT) has therefore emerged as a critical strategy for minimizing radiation burden while maintaining diagnostic performance. Nevertheless, reduced photon flux inevitably results in increased quantum noise, streak artifacts, and loss of low-contrast detectability. This review examines the evolution of LDCT reconstruction from analytical and iterative approaches to modern deep learning reconstruction (DLR) and emerging physics-informed generative models.

Methods

Recent advances in model-based iterative reconstruction, convolutional neural networks, generative adversarial networks, score-based generative models, diffusion probabilistic models, and latent diffusion architectures were systematically reviewed. Particular emphasis was placed on data-consistency constraints, physics-informed learning frameworks, uncertainty quantification, and clinical deployment challenges.

Results

Physics-informed generative reconstruction has demonstrated substantial improvements in image quality, texture preservation, structural fidelity, and sparse-view recovery compared with conventional reconstruction approaches. Diffusion-based models outperform many existing deep learning methods in restoring diagnostically relevant anatomical information while minimizing hallucination risks through incorporation of projectiondomain constraints.

Conclusions

The convergence of physical imaging models and generative artificial intelligence represents a transformative paradigm shift in CT reconstruction. Future reconstruction ecosystems will likely incorporate foundation models, multimodal priors, uncertainty-aware inference, and real-time adaptive optimization, enabling unprecedented reductions in radiation dose without compromising diagnostic accuracy.

Zitieren

APA

Kang, B., Lee, H., Jin, G., & Singh, V. R. (2026). Physics-Informed Generative AI for Low-Dose CT Reconstruction: From Iterative Reconstruction to Diffusion Models. Journal of Medical Imaging, 9(1), 81-93. https://doi.org/10.31916/sjmi2026-01-9

Vancouver

Kang B, Lee H, Jin G, Singh VR. Physics-Informed Generative AI for Low-Dose CT Reconstruction: From Iterative Reconstruction to Diffusion Models. Journal of Medical Imaging. 2026;9(1):81-93. doi:10.31916/sjmi2026-01-9

BibTeX
@article{Kang2026,
  author = {Kang, Byunghoon and Lee, Hwunjae and Jin, Gaehwan and Singh, V. R.},
  title = {Physics-Informed Generative AI for Low-Dose CT Reconstruction: From Iterative Reconstruction to Diffusion Models},
  journal = {Journal of Medical Imaging},
  year = {2026},
  volume = {9},
  number = {1},
  pages = {81--93},
  doi = {10.31916/sjmi2026-01-9}
}
Kostenlos starten →