MAP-Diff anchors diffusion reverse trajectories to clinical intermediate-dose scans, lifting whole-body low-dose PET denoising PSNR from 42.48 dB to 43.71 dB

MAP-Diff treats clinically observed intermediate-dose scans as anchors for the diffusion reverse trajectory and uses timestep-dependent supervision to push the reverse process toward dose-aligned intermediate states. Prior diffusion denoising models achieve strong final reconstructions but their reverse trajectories are typically unconstrained and not aligned with the progressive nature of PET dose formation; this work writes that progression explicitly into the trajectory constraint. The abstract describes the mechanism and reports comparisons on internal and cross-scanner datasets with PSNR, SSIM, and NMAE values relative to 3D DDPM.

Anchor timesteps are calibrated by degradation matching between simulated diffusion corruption and real multi-dose PET pairs, paired with a timestep-weighted anchor loss that stabilizes stage-wise learning. Anchor placement moves from manual specification to data-driven degradation matching against real multi-dose pairs, so supervision lands on stages corresponding to real dose levels. The abstract states the calibration and loss design and credits them with stabilizing stage-wise learning; implementation details and ablations are not expanded in the abstract.

At inference the model needs only ultra-low-dose input and can produce progressive, dose-consistent intermediate restoration. Unlike diffusion methods that target a single final reconstruction, this framework preserves recoverability of intermediate dose states at inference. The abstract explicitly states the inference input requirement and progressive restoration capability, without reporting inference cost or reader-study validation.

On internal and cross-scanner datasets, MAP-Diff consistently outperforms strong CNN-, Transformer-, GAN-, and diffusion-based baselines. The gains are not confined to a single scanner; the cross-scanner cohort also leads, suggesting some device-level generalizability. Internal dataset PSNR 42.48 to 43.71 dB, SSIM 0.986, NMAE 0.115 to 0.103; external cohort 34.42 dB PSNR and 0.141 NMAE, all reported as outperforming every competing method.

The result targets 3D whole-body low-dose PET denoising and suits institutions that have multi-dose paired data available for anchor timestep calibration; because inference needs only ultra-low-dose input, it fits imaging workflows that want progressive, dose-consistent intermediate restoration from a single ultra-low-dose acquisition. The cross-scanner results suggest transferability between Siemens Biograph Vision Quadra and United Imaging uEXPLORER, providing a starting point for testing under multi-center and multi-protocol conditions.

The visible text is abstract-level and omits sample sizes, patient composition, ablations, statistical significance testing, and inference cost, so judgments about robustness of the gains and clinical readability await the full paper. Anchor timestep calibration depends on real multi-dose PET pairs, and how it transfers to settings without such pairs is an open question worth watching. In addition, PSNR, SSIM, and NMAE reflect image fidelity, and their relationship to clinical diagnostic utility is not addressed in the visible text.