Low-dose CT denoising drops the whole-image trade-off: RouteDiff reaches 40.95 dB PSNR on Mayo 2020 within ten network evaluations
Synopsis
Diffusion-based low-dose CT denoising previously applied one structure–detail trade-off to the whole image, set each step's restoration emphasis by the timestep, and derived each next state recursively from the current one, carrying errors forward. RouteDiff routes detail pixel-wise to where a confidence map trusts it, judges the restoration stage from current features and route disagreement through a state hypergraph, and re-enters the trajectory from an anchored estimate at a fixed step. With the same ten network evaluations as its baseline, it achieves the highest PSNR and SSIM on four datasets, 40.95 dB and 95.13% on Mayo 2020 and 45.18 dB and 97.23% on Mayo 2016.
Interpretation
Diffusion samplers for low-dose CT denoising no longer impose a single structure–detail trade-off on the whole image; high-frequency detail is admitted pixel by pixel. Earlier samplers used one global trade-off, set each step's restoration emphasis by the timestep, and derived each next state recursively from the current one, carrying its errors forward. On Mayo 2020, RouteDiff reaches 40.95 dB PSNR and 95.13% SSIM, exceeding the respective runners-up FoundDiff and AMIR by 0.61 dB and 0.31%, and its PSNR is 0.69 dB above CoreDiff.
The restoration stage is judged from the state the network has actually reached rather than from the timestep alone. Prior practice treated the timestep as evidence that all inputs sit at the same degradation level, although within one slice homogeneous tissue may be nearly restored while edges and low-contrast structures are not. In the Mayo 2016 ablation, the timestep-only stage variant falls 0.25 dB below the full model and the variant without route evidence falls 0.11 dB below it.
Sampling re-projects a route-aware anchor onto the trajectory at a fixed step and hands the re-entered state back to be judged again. The recursive update rested every state on its predecessor, so errors already present in the state were carried forward. In the Mayo 2020 design analysis, skipping the re-projection lowers PSNR by 0.39 dB, the second largest drop among the examined variants.
Perspective
The demonstrated setting is 2D slice restoration on four CT datasets with three adjacent slices as input, ten network evaluations at inference and re-entry at step 3; Piglet serves as an external test set using the Mayo 2020 weights directly.
The re-entry step is fixed for all slices rather than chosen from the observed state; restoration is two-dimensional, and extending the state hypergraph to volumetric restoration and assessing outputs in a radiologist reader study remain open directions.
