MR-Guided PET Denoising and Resolution Enhancement Improves Visual Interpretation and Preserves Quantitative Behavior Across Amyloid Tracers
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Amyloid- β PET provides critical biomarker data for Alzheimer’s disease diagnosis and anti-amyloid therapy evaluation, yet low spatial resolution and partial volume effects result in decreased interpretability, particularly in cases with low or borderline cortical amyloid burden. While quantitative metrics (SUVr, Centiloid) aid in interpretation of amyloid burden, disagreement between visual reads and quantitative burden does occur, further blurring the line between positive or negative scans. We evaluated whether a vendor-neutral MR-guided PET denoising and resolution enhancement method (MRG) that uses Bowsher regularization improves image interpretability and reader performance while preserving established quantitative biomarkers across multiple amyloid tracers, leading to increased concordance among visual reads and quantitative metrics.
Methods
Standard (STN) and MRG PET images were compared for four tracers ([ 18 F]AV-45 ([ 18 F]florbetapir, FBP), [ 18 F]florbetaben (FBB), [ 18 F]flutemetamol (FMM), and [ 11 C]Pittsburgh compound-B (PiB)) collectively from 24 MRI and 33 PET scanners. Quantitative equivalence was assessed by comparing Standardized Uptake Value ratio (SUVr) and Centiloid scores. In three of the four tracers (FBP, FBB, FMM), visual-quantitative concordance (AUC) and reader performance were evaluated in a blinded multi-reader study by four highly experienced brain PET readers who assessed image quality, artifact severity, reader confidence, and binary amyloid positivity.
Results
Across all tracers, MRG preserved quantitative SUVr and Centiloid metrics relative to STN ( R 2 > 0.90 for all tracers) without introducing bias to the SUVr metric. Concordance between visual reads and quantitative burden measures significantly improved with MRG. In the multi-reader study, MRG resulted in significantly higher image quality, lower artifact burden, and greater reader confidence compared to STN ( p < 0.0001 ). Reader accuracy increased from 0.89 to 0.94, and the false-negative rate decreased from 0.08 to 0.04. Crucially, improvements in reader confidence, accuracy, and the reduction in false negative reads were most pronounced in cases with low amyloid burden near the threshold of visual positivity.
Conclusions
MRG denoising and resolution enhancement improved perceived image quality, reader confidence, and accuracy for amyloid PET while preserving standard quantitative behavior across tracers. By improving cortical definition in visually challenging low-burden cases without disrupting established SUVr/Centiloid behavior, MRG may reduce visual-quantitative discordance and support more confident amyloid PET interpretation near the threshold of positivity.
