Interpretable deep learning maps spatiotemporal brain aging dynamics aligning with neurodegenerative pathology

Researchers developed a deep learning model to extract spatiotemporal brain aging features from structural MRIs, showing that these deep-learning-derived anatomical changes parallel the progression of tau and amyloid-beta pathology even in cognitively normal individuals.

Researchers trained a 3D convolutional neural network (3D-CNN) on T1-weighted MRIs from 14,555 cognitively normal individuals to estimate biological brain age (BA), achieving a validation mean absolute error (MAE) of 3.04 years.

The model was tested on structural MRI scans from 25,539 cognitively normal UK Biobank participants aged 45–83 years to map biological brain age trajectories and aging dynamics.

Using voxel-wise gradient saliency mapping, the team identified three distinct anatomical foci of biological brain aging: the perihippocampal region (F1), the frontal lobe (F2), and the temporo-parieto-occipital junction (F3).

Traditional morphometrics (cortical thickness, volume, surface area, and curvature) explained on average only 1.6% of the variance in the deep learning model's saliency scores, confirming that deep neural networks capture structural alterations unapparent to standard human measurement.

The spatiotemporal expansion of these deep-learning-salient features was found to mirror the known propagation patterns of tau and amyloid-beta pathology as typically measured by positron emission tomography (PET).

The study challenges the consensus that structural T1-weighted MRI cannot capture anatomical changes associated with early-stage, AD-relevant neuropathology.

Quantifying neurodegenerative anatomy through widely accessible MRI rather than costly, logistically challenging PET scans could democratize patient screening and disease tracking.

Analyzing the first and second derivatives of brain age (velocity and acceleration) showed that biological brain aging accelerates most rapidly between the ages of 60 and 70, providing a precise temporal window for preventative interventions.

This deep learning framework could significantly optimize clinical trial recruitment for neurodegeneration therapeutics by reducing screening costs and failures associated with PET scan eligibility requirements.

Biotechnology companies developing targeted gene or antisense oligonucleotide therapies for Alzheimer's disease can leverage structural MRI-based brain age acceleration and spatial propagation mapping as sensitive exploratory biomarkers.

Longitudinal validation studies to confirm if cross-sectional deviations in MRI-derived saliency accurately predict within-subject cognitive decline and clinical progression over time.

The expansion and validation of the open-source 3D-CNN software on diverse, multi-ethnic patient cohorts to confirm generalizability beyond the UK Biobank dataset.