RouteDiff: Low-Dose CT Denoising via Fidelity-Refinement Routing and Hypergraph State Regulation

Xinyuan Wang, Ruili Li, Tianyang Li, Yan Li, Xuepeng LiuarXiv 2026 · arXiv

TL;DR

RouteDiff is a state-aware diffusion framework for low-dose CT (LDCT) denoising that replaces standard, one-size-fits-all sampling with a dynamic "dual-route" approach. By regulating the restoration process based on the actual image state—rather than just the timestep—and using a hypergraph to group scattered regions in similar stages of recovery, it achieves state-of-the-art results while significantly reducing the risk of "hallucinated" anatomical structures.

What Problem Does This Paper Address?

Computed Tomography (CT) is a vital diagnostic tool, but ionizing radiation carries cumulative health risks. Reducing the "dose" (tube current) introduces heavy noise and streak artifacts that obscure fine details.

While diffusion models have emerged as powerful tools for restoring these images, they suffer from three practical flaws:

  1. Global Trade-offs: They often apply the same balance of "structure vs. detail" to the entire image, even though some regions (like homogeneous tissue) need less refinement than others (like complex bone structures).
  2. Timestep Blindness: Standard samplers assume every pixel reaches the same level of restoration at the same timestep. In reality, different anatomies restore at different rates.
  3. Recursive Error Accumulation: Errors in early sampling steps are carried forward and amplified, leading to "hallucinations"—plausible-looking but fake anatomical details.

Methods

RouteDiff introduces three novel modules to a shared U-Net backbone:

  • Fidelity–Refinement Routing (FRR): The model predicts two versions of the image: a conservative "fidelity" route (safe structure) and a detail-oriented "refinement" route. A rule-based confidence map fuses them pixel-wise, only admitting high-frequency details from the refinement route where they can be trusted. The disagreement between these two routes is treated as "evidence" of the current restoration state.
  • Hypergraph State Regulation (HSR): Instead of relying solely on the timestep $t$, HSR builds a "state hypergraph." Nodes represent image positions, and "soft hyperedges" group regions that are in a similar state of restoration (e.g., grouping scattered lung regions together). This hypergraph modulates the network's bottleneck and skip connections, allowing the model to adapt its behavior to the actual progress of the image.
  • Estimate-Anchored Re-entry (EAR): To break the cycle of error accumulation, the model performs a mid-sampling "reset." At a fixed step, it constructs a route-aware anchor (a clean estimate), re-projects it onto the diffusion trajectory, and restarts the sampling from this corrected state.

Datasets

The authors evaluated RouteDiff across four diverse datasets:

  • Mayo 2020: Chest and abdominal scans with validated projection-domain noise insertion.
  • Mayo 2016: The classic AAPM Low Dose CT Grand Challenge dataset.
  • Piglet: A real-world multi-dose dataset used as an external test set to evaluate generalization to unseen doses and scanner noise.
  • AMOS: A large-scale abdominal benchmark used to test how denoising affects downstream organ segmentation (using a frozen nnU-Net).

Key Insights and Ideas

  • State > Timestep: The core "aha moment" is that the network should know how much it has restored, not just how long it has been running. By judging the stage from bottleneck features and route disagreement, the model avoids over-smoothing or over-hallucinating.
  • Hypergraphs for Non-Local Context: Traditional CNNs are local, and Transformers are pairwise. Hypergraphs allow the model to group arbitrary, scattered regions (like both kidneys or separated lung lobes) into a single "state," ensuring consistent restoration across similar tissues.
  • Corrective Re-entry: The EAR module acts like a "sanity check." By re-anchoring the estimate mid-way, the model sheds accumulated sampling noise, leading to much sharper and more faithful textures compared to purely recursive samplers.

Key Results

RouteDiff achieved state-of-the-art performance across all benchmarks using only 10 network evaluations (NFE):

  • Mayo 2020: Reached 40.95 dB PSNR, outperforming the strong CoreDiff baseline by 0.69 dB.
  • Mayo 2016: Achieved 45.18 dB PSNR and 97.23% SSIM, significantly higher than competitors like RED-CNN and FoundDiff.
  • Generalization: On the Piglet dataset, RouteDiff maintained high performance on unseen doses (1/2 and 1/20 dose), showing it doesn't just memorize training noise patterns.
  • Clinical Utility: In the AMOS segmentation task, RouteDiff-restored images yielded the highest Dice score (80.92%), proving that its denoising actually helps automated diagnostic tools perform better.

Limitations and Future Work

  • Fixed Re-entry: Currently, the re-entry step ($m$) is fixed. The authors suggest that selecting this step adaptively based on the HSR stage judgment could further improve results.
  • 2D vs. 3D: The model processes 2D slices with a three-slice context. Extending the state hypergraph to 3D volumetric restoration is a logical next step for better inter-slice consistency.
  • Radiologist Validation: While quantitative metrics (PSNR/SSIM) and segmentation scores are high, the authors note that a formal reader study by radiologists is needed to confirm the clinical "trustworthiness" of the restored details.

Summary updated

Abstract

Diffusion models have become a strong prior for low-dose computed tomography (LDCT) restoration. However, existing samplers apply one structure-detail trade-off to the whole image, set the restoration emphasis of each step by the timestep, and derive each next state recursively from the current one, carrying its errors forward. We propose RouteDiff, a state-aware dual-route diffusion framework that routes detail to where it can be trusted, regulates each reverse step by the state it has reached, and re-enters the trajectory from an anchored estimate. Fidelity-Refinement Routing (FRR) equips one shared U-Net with a conservative fidelity route and a detail-oriented refinement route, fuses them pixel-wise with a confidence map, and exposes their disagreement as route evidence. Hypergraph State Regulation (HSR) organizes the current features and this evidence into a state hypergraph, whose soft hyperedges group scattered regions in the same state. Its readout yields an ordered stage judgment that modulates the bottleneck, and its incidence guides feature interaction in the skip connections. Estimate-Anchored Re-entry (EAR) re-projects a route-aware anchor onto the trajectory at a fixed step and hands the new state back to HSR to be judged again. With the ten network evaluations of its baseline, RouteDiff achieves the best peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) on four datasets with inserted, real and simulated noise (40.95 dB and 95.13% on Mayo 2020; 45.18 dB and 97.23% on Mayo 2016). It also generalizes to unseen doses of real scans and yields the highest Dice (80.92%) in downstream segmentation.
BibTeX
@misc{wang2026routediff,
  author = {Xinyuan Wang and Ruili Li and Tianyang Li and Yan Li and Xuepeng Liu},
  title = {{RouteDiff: Low-Dose CT Denoising via Fidelity-Refinement Routing and Hypergraph State Regulation}},
  year = {2026},
  eprint = {2611.18786},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18786},
}
Figure