Biomarkers.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Magnetic Resonance Imaging (MRI) is vital for Alzheimer's disease (AD) diagnosis and monitoring.
• Data variability across national cohorts challenges consistent AD diagnosis using MRI.
• Explainable AI with attention maps can enhance diagnostic accuracy and interpretability in AD through MRI analysis.

Purpose of the Study:

• To develop a deep learning approach for visualizing pathological brain regions in Alzheimer's disease (AD) versus cognitively normal (CN) classification.
• To enhance diagnostic accuracy and interpretability of MRI data across multi-cohort datasets for AD detection.

Main Methods:

• Utilized data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Gwangju Alzheimer's and Related Dementia (GARD) databases.
• Trained a ResNet-50 classification model with an attention mechanism to differentiate AD from CN individuals.
• Visualized attention maps in sagittal, axial, and coronal planes of MRI scans to identify key pathological regions.

Main Results:

• Achieved high performance in distinguishing AD from CN individuals: 98.18% accuracy, 96.73% specificity, and 98.97% sensitivity.
• Attention map analysis highlighted the hippocampus and temporal lobe as primary regions of interest, aligning with known AD pathology.
• Demonstrated consistent performance on the GARD cohort, confirming the model's generalizability.

Conclusions:

• The proposed method excels in AD vs. CN classification, accurately identifying critical pathological areas like the hippocampus and temporal lobe.
• Consistent results across diverse cohorts underscore the robustness and potential for clinical implementation of this explainable AI approach.