Biomarkers.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Neuropathology

Background:

• Clinical dementia diagnosis faces accuracy challenges, with neuropathology as the gold standard.
• Deep learning models often trained on clinical diagnoses may limit dementia type prediction accuracy.
• This study focuses on quantifying atrophy for Alzheimer's, Vascular, and Lewy Body Dementias.

Purpose of the Study:

• Develop a multi-label deep learning network trained on neuropathologically confirmed diagnoses.
• Quantify brain atrophy patterns specific to Alzheimer's Disease (AD), Vascular Dementia (VD), and Lewy Body Dementia (LBD).
• Create novel, noninvasive neuroimaging metrics for pathology identification.

Main Methods:

• Trained deep learning models on 3D T1-weighted MRI scans from 423 demented participants and 361 controls.
• Generated explainable heatmaps to visualize pathology-specific patterns.
• Developed Deep Signature of Pathology Atrophy REcognition (DeepSPARE) indices to quantify brain alterations.

Main Results:

• Achieved balanced accuracies of 0.844 (AD), 0.839 (VD), and 0.623 (LBD) via 5-fold cross-validation.
• Explainable heatmaps highlighted distinct brain regions for AD (hippocampus), VD (white matter), and LBD (occipital).
• DeepSPARE indices correlated significantly with pathology-specific measures (cognition, Braak stages, white matter hyperintensity).

Conclusions:

• Demonstrated a deep learning framework for identifying antemortem neuroimaging signatures of distinct dementia pathologies.
• The novel DeepSPARE indices are accurate, pathology-specific, and noninvasive.
• Connected in-vivo T1 imaging with pathological identifications for improved dementia diagnosis.