AmyloidPETNet: Classification of Amyloid Positivity in Brain PET Imaging Using End-to-End Deep Learning.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Neuroimaging Analysis
• Machine Learning for Diagnostics

Background:

• Visual assessment of amyloid PET scans requires radiologist expertise.
• Quantification of amyloid burden often necessitates computationally expensive MRI processing.
• Developing automated methods can enhance accessibility and efficiency in amyloid PET analysis.

Purpose of the Study:

• To develop and evaluate a deep learning model (AmyloidPETNet) for classifying amyloid PET scans.
• To assess the model's performance on independent datasets and various tracers.
• To compare the model's accuracy against human visual reads.

Main Methods:

• Trained AmyloidPETNet on 1538 PET scans (766 patients) and validated on 205 scans (95 patients).
• Tested the model on independent datasets including ADNI, OASIS, and A4 study scans with different tracers.
• Compared model performance using AUC, other metrics, and Cohen κ for physician-model agreement.

Main Results:

• AmyloidPETNet achieved high AUCs (≥0.95) across multiple datasets and tracers, demonstrating strong generalization.
• The model showed excellent performance on ADNI ¹⁸F-FBP scans (AUC=0.97) and generalized to other tracers (AUC≥0.97).
• Physician-model agreement varied from fair (Cohen κ=0.39) to almost perfect (Cohen κ=0.93), indicating robust performance.

Conclusions:

• The developed deep learning model accurately classifies amyloid PET scans as positive or negative.
• AmyloidPETNet offers an automated solution, reducing the need for expert readers and structural MRI.
• This AI-driven approach has the potential to improve the efficiency and accessibility of amyloid PET diagnostics.