Learning to synthesise the ageing brain without longitudinal data.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Predicting future brain appearance and disease progression is challenging due to the difficulty of collecting longitudinal data.
• Understanding age-related brain changes and the impact of neurodegenerative diseases like Alzheimer's Disease (AD) is crucial for early detection and intervention.

Purpose of the Study:

• To develop a deep learning method for simulating subject-specific brain aging trajectories using only cross-sectional data.
• To synthesize brain images that reflect both chronological age and Alzheimer's Disease status.
• To preserve subject identity during the image synthesis process.

Main Methods:

• A deep learning model employing an adversarial formulation to learn the joint distribution of brain appearance, age, and AD status.
• Reconstruction losses were defined to maintain subject identity.
• The model was trained and evaluated on two widely used datasets, comparing against several benchmarks.

Main Results:

• The model successfully synthesized realistic brain images conditioned on age and AD status.
• Despite using cross-sectional data, the model identified gray matter atrophy patterns in the middle temporal gyrus characteristic of AD.
• The model demonstrated generalization ability by performing well when trained on one dataset and tested on another.

Conclusions:

• The proposed method effectively separates the influences of aging, Alzheimer's Disease, and individual anatomy from 2D cross-sectional brain data.
• This approach holds significant potential for large-scale neurodegenerative disease studies, especially those integrating diverse data sources.
• The open-source code facilitates community adoption and further research in the field.