Deep learning-based enhancement of ultra-low-dose multi-tracer PET/MR: a technical feasibility study in Alzheimer’s disease

Abstract Background To develop a Dual-Branch Dense Wavelet Attention Network (DDWAN) for enhancing simulated ultra-low-dose multi-tracer (Aβ, tau and FDG) PET imaging acquired on a PET/MR system and to evaluate its technical feasibility in Alzheimer’s disease (AD), using full-dose images acquired under the same protocol as references. Methods This retrospective study included patients underwent 18 F-florbetapir (Aβ), 18 F-APN-1607 (tau), or 18 F-FDG (N) PET/MR. DDWAN was trained to generate high-quality images from simulated ultra-low-dose PET data (1/100–1/5 of standard dose) with co-registered T1WI as anatomical guidance. Image quality was quantified quantitatively via Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). Quantitative agreement was evaluated by comparing Standardized Uptake Value Ratios (SUVR) and Centiloid values between DDWAN-enhanced low-dose and standard-dose images. Qualitative assessment was performed by nuclear medicine physicians using a 5-point Likert scale. Diagnostic performance was evaluated using standard-dose visual interpretation as the reference standard. Results The DDWAN model significantly enhanced image quality across all tracers and dose levels. Quantitative metrics demonstrated substantial improvements, with PSNR increasing most notably at the 1/100 dose (ΔPSNR: +7.11 dB for Aβ, + 11.75 dB for tau, + 4.82 dB for FDG; all Bonferroni-corrected P < 0.0125). Excellent quantitative agreement was observed, with minimal bias in SUVR and Centiloid values between enhanced low-dose and standard-dose images (all P >0.05). Qualitative scores for DDWAN-enhanced images were significantly superior to original low-dose images and comparable to standard-dose references, with the majority of 1/20 dose images rated as diagnostically acceptable (score ≥ 3). Diagnostic agreement remained high across all tracers, with sensitivity and specificity exceeding 90% even at the 1/100 dose level for 18 F-APN-1607 and 18 F-FDG. Conclusion DDWAN improved image quality and quantitative consistency of simulated ultra-low-dose multi-tracer (Aβ, tau and FDG) PET/MR images compared with full-dose references acquired under the same protocol. These results support technical feasibility, while further validation using standard tracer-specific acquisition windows is required before clinical application.