MRI-free amyloid PET quantification using a deep learning model and white matter reference

Abstract

Background Accurate quantification of standardized uptake value ratio (SUVR) in amyloid PET is essential for Alzheimer's disease (AD) diagnosis but typically requires MRI-based segmentation due to subtle uptake differences between gray and white matter. Objective This study aimed to develop and validate a 3-dimensional deep learning model capable of segmenting these tissues directly from PET images to enable MRI-free SUVR quantification for AD diagnosis. Methods This retrospective study included 385 participants who underwent brain amyloid PET and MRI. After excluding 12 data-corrupted cases, 373 subjects were divided into training (n = 318) and test (n = 55) sets. External validation used 625 PET/CT scans from the Alzheimer's Disease Neuroimaging Initiative. Model performance was assessed using Dice coefficients and intersection over union. PET-based SUVRs derived from model-generated masks were compared with MRI-based SUVRs using Spearman correlation, and their diagnostic utility was evaluated by group differences and receiver operating characteristic analysis. Results The model achieved high Dice coefficients for grey matter (GM; 0.785 internal, 0.743 external) and white matter (WM; 0.838 internal, 0.803 external). PET/CT-based SUVR values strongly correlated with MRI references (Spearman's ρ ≥ 0.98, p < 0.001). PET/CT-derived SUVRGM and SUVRGM/WM predicted amyloid status (AUC 0.86 and 0.85, respectively) and cognitive impairment (AUC 0.78). Conclusions Deep learning–based amyloid PET segmentation enables accurate MRI-free quantification of gray and white matter SUVRs. This approach simplifies clinical workflow while maintaining diagnostic performance comparable to MRI-based methods for AD diagnosis.