γ-Photon Flow-Field Denoising Gets Structure Guidance at Every Layer and Timestep-Tuned Frequency Boost, Reaching 28.325 dB PSNR

Denoising of γ-photon flow-field images no longer relies on a single concatenation of the degraded image at the input; the structural prior is reinjected at every layer of reverse denoising. A conventional conditional diffusion model introduces the condition image by channel concatenation, and that structural information is diluted across downsampling stages, so it cannot persistently constrain the main flow contour and boundary positions. On the 30 image pairs of the phantom-level independent test set, which received no cropping or rotation augmentation, the full SFGD reaches 28.325 dB PSNR and 0.926 SSIM, against 22.826 dB and 0.773 for the conditional DDPM baseline.

The strength and range of frequency enhancement change with reverse sampling, amplifying little noise early and restoring edge detail later. Existing frequency-domain methods set modulation parameters from fixed transforms or current input content rather than explicitly from the diffusion timestep, so they do not track the changing noise level. In the ablation, the TAFM-only configuration raises PSNR from 22.826 dB to 24.357 dB and SSIM from 0.773 to 0.895; the SCGM-only configuration reaches 26.590 dB and 0.862; the two together reach 28.325 dB and 0.926.

Structure guidance comes from a multi-scale structure stream aligned to the diffusion-branch resolution and fused layer by layer through windowed cross-attention. That stream uses parallel 3×3 and 7×7 depthwise separable convolution paths for local edges and larger-range structural orientation, combined by spatially gated fusion. Cross-attention is computed inside non-overlapping 4×4 windows, and a residual scaling coefficient initialized to zero makes the module approximate the original diffusion branch early in training; skip connections are split into separate noise-branch and structure-enhanced paths.

The training objective constrains noise, image structure and spectral amplitude together, with the three weights set automatically from task uncertainty. The three losses differ in scale and optimization difficulty, so fixed weights would require repeated hyperparameter search and could not follow changes in relative difficulty during training. The structural-consistency loss is computed only at timesteps satisfying t<250, because the approximate clean image inferred from predicted noise is unreliable at high-noise steps; the frequency-consistency loss constrains amplitude only and leaves phase unconstrained.

The demonstrated scope is γ-photon flow-field images generated by GATE Monte Carlo simulation: 97 phantoms at 128×128 pixels reconstructed with OSEM, with training and validation expanded by four-quadrant cropping and right-angle rotation, and an independent test set of 30 original pairs without augmentation. The intended readers are researchers and engineers working on photon-imaging reconstruction and flow-field image preprocessing.

All quantitative results come from GATE-simulated and OSEM-reconstructed phantom images, with no measured γ-photon flow-field data. Metrics for every method are lower on the oblique-wake group than on the other two groups, which the source attributes to more diffuse contour transitions and fewer usable structural cues in the input. The data-availability statement says datasets are available from the corresponding author on reasonable request.