MAP-Diff: Multi-Anchor Guided Diffusion for Progressive 3D Whole-Body Low-Dose PET Denoising

Peiyuan Jing; Chun-Wun Cheng; Liutao Yang; Zhenxuan Zhang; Thiago V. Lima; Klaus Strobel; Antoine Leimgruber; Angelica Aviles-Rivero

cs.CVMar 2, 2026

MAP-Diff: Multi-Anchor Guided Diffusion for Progressive 3D Whole-Body Low-Dose PET Denoising

Authors: Peiyuan Jing, Chun-Wun Cheng, Liutao Yang, Zhenxuan Zhang, Thiago V. Lima, Klaus Strobel, Antoine Leimgruber, Angelica Aviles-Rivero,

Organizations: School of Engineering, Zurich University of Applied Sciences, CH · Bioengineering Department and Imperial-X, Imperial College London, UK · DAMTP, University of Cambridge, UK · Lucerne University Teaching and Research Hospital, CH ·

Abstract

Low-dose Positron Emission Tomography (PET) reduces radiation exposure but suffers from severe noise and quantitative degradation. Diffusion-based denoising models achieve strong final reconstructions, yet their reverse trajectories are typically unconstrained and not aligned with the progressive nature of PET dose formation. We propose MAP-Diff, a multi-anchor guided diffusion framework for progressive 3D whole-body PET denoising. MAP-Diff introduces clinically observed intermediate-dose scans as trajectory anchors and enforces timestep-dependent supervision to regularize the reverse process toward dose-aligned intermediate states. Anchor timesteps are calibrated via degradation matching between simulated diffusion corruption and real multi-dose PET pairs, and a timestep-weighted anchor loss stabilizes stage-wise learning. At inference, the model requires only ultra-low-dose input while enabling progressive, dose-consistent intermediate restoration. Experiments on internal (Siemens Biograph Vision Quadra) and cross-scanner (United Imaging uEXPLORER) datasets show consistent improvements over strong CNN-, Transformer-, GAN-, and diffusion-based baselines. On the internal dataset, MAP-Diff improves PSNR from 42.48 dB to 43.71 dB (+1.23 dB), increases SSIM to 0.986, and reduces NMAE from 0.115 to 0.103 (-0.012) compared to 3D DDPM. Performance gains generalize across scanners, achieving 34.42 dB PSNR and 0.141 NMAE on the external cohort, outperforming all competing methods.

Figures & tables

Figure 1: Overview of MAP-Diff. (a) Conceptual comparison between standard diffusion training with end-point supervision and the proposed multi-anchor trajectory-level supervision. (b) MAP-Diff training strategy with timestep zoning and anchor-weighted loss. (c) Progressive inference trajectory enabling dose-consistent intermediate outputs.

Method	Internal (University Hospital of Bern)			External (Shanghai Ruijin Hospital)
	PSNR $\uparrow$	SSIM $\uparrow$	NMAE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	NMAE $\downarrow$
Low-dose Input	38.545 $\pm$ 1.798**	0.948 $\pm$ 0.018**	0.199 $\pm$ 0.036**	30.402 $\pm$ 3.475**	0.896 $\pm$ 0.033**	0.257 $\pm$ 0.059**
3D cGAN ( Wang et al., 2018 )	40.721 $\pm$ 1.293**	0.977 $\pm$ 0.006**	0.127 $\pm$ 0.015**	33.062 $\pm$ 3.624**	0.927 $\pm$ 0.053**	0.174 $\pm$ 0.031**
3D UNet ( Çiçek et al., 2016 )	41.810 $\pm$ 1.377**	0.972 $\pm$ 0.007**	0.120 $\pm$ 0.015**	33.654 $\pm$ 3.810**	0.904 $\pm$ 0.074**	0.164 $\pm$ 0.031**
RED-CNN ( Chen et al., 2017 )	42.185 $\pm$ 1.427**	0.981 $\pm$ 0.005**	0.119 $\pm$ 0.016**	33.989 $\pm$ 3.725**	0.945 $\pm$ 0.025**	0.161 $\pm$ 0.030**
EDCNN ( Liang et al., 2020 )	42.355 $\pm$ 1.579**	0.983 $\pm$ 0.006**	0.116 $\pm$ 0.017**	33.841 $\pm$ 3.615**	0.942 $\pm$ 0.023**	0.164 $\pm$ 0.024**

Table 1: Quantitative comparison results. Statistical significance is assessed using a two-sided Wilcoxon signed-rank test with Holm correction. * and ** denote statistically significant differences compared with MAP-Diff at

p<0.05

and

p<0.01

, respectively.

Figure 2: Visual comparison of PET denoising across methods.

Figure 3: Intermediate-dose reconstruction accuracy at anchor timesteps. (a) Quantitative comparison against paired real-dose references at the 1/10, 1/4, and 1/2 dose levels. (b) Coronal-view signed error maps against paired real-dose references.

Baseline 3D DDPM ( Yu et al., 2025 ) (No Anchors): PSNR: 42.483, SSIM: 0.983, NMAE: 0.115
Set	#A	Anchors	W	PSNR	SSIM	NMAE	Set	#A	Anchors	W	PSNR	SSIM	NMAE
S1	3	{1/10,1/4,1/2}	Poly	43.710	0.986	0.103	S2	2	{1/10,1/4}	Poly	43.672	0.984	0.104
S3	2	{1/10,1/2}	Poly	43.521	0.983	0.106	S4	2	{1/4,1/2}	Poly	43.414	0.982	0.107
S5	1	{1/10}	Poly	43.503	0.983	0.106	S6	1	{1/4}	Poly	43.267	0.982	0.109
S7	1	{1/2}	Poly	42.790	0.981	0.111	S8	3	{1/10,1/4,1/2}	Const	41.856	0.905	0.133

Table 2: Ablation study of anchor supervision design in MAP-Diff, evaluating the number of anchors and timestep-weighted loss. Best results in bold.

Image Denoising

Image denoising removes unwanted noise while preserving the edges, textures, and colors needed for faithful images.

See all 157 papers

AnF-DiffPET: Anatomy- and Frequency-Guided Diffusion for PET/CT Denoising
Jul 1, 2026Xuepeng Liu, Ruili Li, Zetong Liu +5Diffusion Models for Image Restoration Diffusion Model Guidance
College of Medicine and Biological Information Engineering, Northeastern University, Shenyang 110169, China · Tohoku University School of Medicine, Sendai 980-8575, Japan · Department of Clinical Neurosciences, University of Cambridge, UK
Less Is More: Training-Free Acceleration Framework of 3D Diffusion Models for Low-Count PET Denoising via Global-Local Trajectory Reduction
Jun 7, 2026Yuhan Liu, Scott M. Leonard, Marlee Crews +5Diffusion Models for Image Restoration Diffusion Model Caching
Department of Radiology, Northwestern University, Chicago, IL, USA
Efficient 3D Whole-Body PET Image Denoising via Conditional Rectified Flow With Optimized Sampling Strategy
Sep 15, 2026Jiale Shen, Guolin Wang, Chenhao Wang +3Rectified Flow 3D Medical Imaging
College of Biomedical Engineering & Instrument Science, Zhejiang University, Zhejiang, 310063, China

Jul 1, 2026J/K move · Enter open · S save

AnF-DiffPET: Anatomy- and Frequency-Guided Diffusion for PET/CT Denoising

Xuepeng Liu, Ruili Li, Zetong Liu, Renyiming Li+4

College of Medicine and Biological Information Engineering, Northeastern University, Shenyang 110169, China · Tohoku University School of Medicine, Sendai 980-8575, Japan · Department of Clinical Neurosciences, University of Cambridge, UK

Diffusion Models for Image Restoration Diffusion Model Guidance PET/CT Imaging Image Denoising

Open paper AskPDF arXiv

Positron emission tomography (PET) provides essential functional information for disease assessment, however reducing injected activity or acquisition time produces low-dose (LD) PET with stronger count dependent noise and less reliable uptake quantification. Diffusion models offer a promising solution for PET denoising by progressively recovering high-dose (HD) PET images from LD inputs. However, LD-to-HD PET denoising is still challenging due to insufficient anatomical guidance, unstable multi-scale feature propagation, and uncertain frequency domain uptake recovery. We propose AnF-DiffPET, an anatomy- and frequency-guided diffusion framework for computed tomography (CT) conditioned LD PET denoising. The framework integrates Anatomical-Frequency Guidance (AFG), Multi-Scale Cross-Transformer Reconstruction (MSCTR), and Frequency-Contrastive Hard Mining (FCHM) to enhance anatomy aware feature modulation and frequency domain consistency during denoising. Experimental results across four PET/CT datasets show that the proposed method improves image fidelity, anatomical consistency, and quantitative fidelity over representative CNN-based, GAN-based, transformer-based, and diffusion-based methods. The code and trained models will be publicly released upon acceptance.

Abstract

Image Denoising

Explore similar work

AnF-DiffPET: Anatomy- and Frequency-Guided Diffusion for PET/CT Denoising

Less Is More: Training-Free Acceleration Framework of 3D Diffusion Models for Low-Count PET Denoising via Global-Local Trajectory Reduction

Efficient 3D Whole-Body PET Image Denoising via Conditional Rectified Flow With Optimized Sampling Strategy