Generalizability of diffusion-based CT reconstruction across scanners and protocols

Investigate the generalizability of diffusion-based computed tomography reconstruction methods across different scanners, acquisition geometries, and acquisition protocols to ascertain how well such models transfer to diverse real-world CT settings.

Background

The paper introduces DM4CT, a comprehensive benchmark evaluating diffusion models for CT reconstruction across medical, industrial, and synchrotron datasets. Despite strong performance in controlled scenarios, the authors highlight practical challenges such as distribution shifts and varying acquisition conditions that may hinder deployment.

Recognizing that CT systems vary widely in scanner hardware, geometry (e.g., parallel-beam, cone-beam, helical), and acquisition protocols, the authors note that it remains uncertain how well diffusion-based reconstruction methods trained under one setting perform under mismatched conditions. They suggest extending DM4CT with multi-institutional or cross-protocol datasets to rigorously evaluate transferability.

References

Finally, a key open question is the generalizability of diffusion-based reconstruction across scanners, geometries, and acquisition protocols. Extending DM4CT with multi-institutional or cross-protocol datasets would enable rigorous testing of how well these models transfer to diverse real-world CT settings.

— DM4CT: Benchmarking Diffusion Models for Computed Tomography Reconstruction  (2602.18589 - Shi et al., 20 Feb 2026) in Conclusion, Future Work

Due to limitations in accessing clinical CT datasets, we were unable to conduct comprehensive evaluations across diverse real medical use cases. Instead, we have validated our method on synthetic human-body datasets containing metal implants (e.g., in soft tissue and dental regions). Additionally, we have tested on a real-world dataset where metallic wires were inserted into a chicken to induce metal artifacts in biological tissues. These experiments show that our method generalizes well to various anatomical objects and metals. We are actively seeking access to clinical data to further investigate the generalizability of our method under real-world scenarios for future work.

— Splat-based Metal Artifact Reduction in Cone-Beam CT via Polychromatic Modeling  (2608.13159 - Choi et al., 13 Aug 2026) in Section Discussion, Sec. 6