DeepAD

DeepAD predicts amyloid deposition from ^18F-florbetapir PET scans using deep learning to estimate standardized uptake value ratio (SUVR) for Alzheimer's disease diagnosis and prognosis.

Key Features:

• Deep Learning Architecture: Employs convolutional neural networks (CNNs) and Gradient Boosting Decision Tree algorithms, evaluating ResNet, EfficientNet, and RegNet families and selecting RegNet X064 for best performance.
• Data Integration: Integrates ^18F-florbetapir PET imaging with clinical and genetic data from 2,980 patients in the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort.
• Model Optimization: Optimizes SUVR regression performance via grid search tuning and uses three axial input slices combined with clinical and genetic features.
• Performance Metrics: Reports a mean absolute error (MAE) of 0.0441 and 96.4% accuracy on a 596-patient test set for SUVR prediction.
• Computational Efficiency: Reduces computation time compared to conventional approaches.

Scientific Applications:

• Alzheimer's diagnosis and prognosis: Provides SUVR estimates from ^18F-florbetapir PET for assessment of cortical amyloid burden to support Alzheimer's disease diagnosis and prognosis.
• Patient stratification and personalized medicine: Combines imaging, clinical, and genetic data to inform patient stratification and individualized assessment.
• Extension to other imaging tasks: Potentially adaptable to other neuroimaging tasks requiring quantitative PET-based biomarker prediction.

Methodology:

Uses CNNs and Gradient Boosting Decision Tree algorithms, evaluating ResNet, EfficientNet, and RegNet architectures (selecting RegNet X064); integrates ^18F-florbetapir PET with clinical and genetic data from 2,980 ADNI patients; employs three axial input slices; optimizes models via grid search for SUVR regression and evaluates performance on a 596-patient test set (MAE 0.0441).