Amyloid PET quantification with deep learning segmentation models without MRI.

Abstract

PURPOSE: To develop and validate deep learning (DL) segmentation models for accurate, MRI-less amyloid PET quantification through the Centiloid scale. METHODS: In this retrospective study, two DL models (PET/CT and PET-only) were trained using 2774 18F-FBP PET/MRI scan pairs from the ADNI dataset, where synthetic CT were generated from T1w MRI and used as training data. External validation was performed on 424 PET/CT scans from the OASIS-3 dataset. The performance of the DL models was compared to the MRI-based FreeSurfer reference method and the template-based rPOP method. Statistical analyses included the calculation of mean absolute error (MAE), R-squared (R²), equivalence testing, and diagnostic accuracy for amyloid positivity (Centiloid > 20), with a focus on the 0–40 Centiloid ‘gray zone’. RESULTS: The PET/CT model demonstrated the highest accuracy, with a MAE of 5.79 Centiloid and a strong correlation (R² = 0.96). The PET-only and rPOP methods showed higher errors with MAEs of 10.8 and 14.45 Centiloid, respectively. The PET/CT model’s Centiloid values were shown to be equivalent to the MRI-derived values within a ± 5 margin, while the PET-only and rPOP methods were not. In the Centiloid gray zone subgroup (n = 93), the PET/CT model yielded the highest diagnostic accuracy of 90.3% and agreement (Cohen’s kappa = 0.79). CONCLUSION: DL models, particularly the PET/CT model trained with synthetic CT, provide accurate and robust Centiloid quantification for 18F-FBP amyloid PET, suggesting their potential as a tool for MRI-less amyloid PET quantification.