Forecasting Future Anatomies: Longitudinal Brain Mri-to-Mri Prediction.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Predicting future brain states from MRI is crucial for understanding neurodegenerative diseases like Alzheimer's disease (AD).
• Current methods often focus on cognitive scores, not direct image prediction.
• Longitudinal MRI analysis is key to modeling disease progression.

Purpose of the Study:

• To investigate deep learning for longitudinal MRI image-to-image prediction.
• To forecast entire brain MRIs several years into the future.
• To model complex, spatially distributed neurodegenerative patterns.

Main Methods:

• Implemented and evaluated five deep learning architectures: UNet, U2-Net, UNETR, Time-Embedding UNet, and ODE-UNet.
• Utilized two longitudinal cohorts: Alzheimer's Disease Neuroimaging Initiative (ADNI) and Australian Imaging Biomarkers and Lifestyle (AIBL).
• Compared predicted follow-up MRIs with actual scans using global similarity and local difference metrics.

Main Results:

• The best-performing deep learning models achieved high-fidelity MRI predictions.
• All evaluated models demonstrated robust cross-cohort generalization to an independent dataset.
• Deep learning reliably predicted participant-specific brain MRIs at the voxel level.

Conclusions:

• Deep learning enables accurate, voxel-level prediction of future brain MRIs.
• This approach offers a novel method for individualized prognosis in neurodegenerative diseases.
• Future research can leverage these techniques for enhanced disease monitoring and treatment strategies.