Alzheimer's Imaging Consortium.

Area of Science:

• Neuroimaging
• Machine Learning
• Alzheimer's Disease Research

Background:

• Alzheimer's disease (AD) diagnosis and prognosis are challenging due to clinical variability.
• Predicting clinical scores like ADAS from MRI is less explored but crucial for assessing severity and aiding prognosis.
• Limited labeled data in AD research hinders deep learning model training.

Purpose of the Study:

• To investigate the efficacy of fine-tuning a pretrained brain age prediction model for predicting Alzheimer's Disease Assessment Scale (ADAS) scores.
• To address the challenge of limited labeled data in medical imaging for Alzheimer's disease research.
• To enhance the prediction of clinical severity and aid in prognosis using MRI-based deep learning models.

Main Methods:

• Developed an ensemble (n=5) model for brain age prediction from 3D brain MRI, employing robust preprocessing, data augmentation, and regularization for generalizability.
• Utilized 11,041 MRIs from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, splitting into training, validation, and test sets.
• Fine-tuned the brain age model to predict ADAS13 scores and evaluated its performance on the validation and test sets.

Main Results:

• Achieved a Mean Absolute Error (MAE) of 5.66, 6.46, and 5.90 for ADAS13 prediction on the training, validation, and test sets, respectively.
• Obtained an R² score of 0.58 (r=0.76, p<<0.01) on the test set, indicating strong predictive performance.
• Demonstrated robust generalization to the test set using only 50% of the available training data.

Conclusions:

• The fine-tuned brain age model effectively predicts ADAS13 scores, demonstrating robustness and generalizability.
• This approach requires less data, outperforming previous methods and offering a solution for training deep learning models with limited medical imaging datasets.
• The study paves the way for developing more effective diagnostic and prognostic tools for Alzheimer's disease.