Fast and Accurate Amyloid Brain PET Quantification Without MRI Using Deep Neural Networks

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Amyloid PET imaging is crucial for diagnosing Alzheimer's disease.
• Current quantification methods often rely on co-registered MRI or CT scans.
• There is a need for non-invasive, MRI-independent quantification techniques.

Purpose of the Study:

• To develop and validate a deep learning-based spatial normalization (SN) method for amyloid PET quantification.
• To assess the accuracy of this method without requiring structural MRI or CT scans.
• To compare the performance of the deep learning SN method against traditional SPM-based SN.

Main Methods:

• A deep neural network was trained on 994 multicenter amyloid PET images and corresponding MRI scans.
• The method was evaluated against FreeSurfer-based quantification using 148 additional PET images.
• External validation was performed on an independent dataset of 136 amyloid PET scans.
• SPM-based SN using MRI was used as a comparative method.

Main Results:

• The deep learning-based SN method demonstrated stronger correlations with FreeSurfer estimates (R²=0.986) compared to SPM-based SN (R²=0.946).
• The proposed method showed superior performance in external validation, outperforming SPM SN without MRI.
• Quantification results showed high agreement across different amyloid PET radiotracers (18F-flutemetamol, 18F-florbetaben, 18F-florbetapir).

Conclusions:

• A novel deep learning-based SN method enables accurate quantitative analysis of amyloid PET images without structural MRI.
• This MRI-independent approach shows high concordance with established MRI-parcellation-based methods.
• The proposed method offers a valuable tool for Alzheimer's disease research and diagnostics using amyloid PET.