Comparison of Explainable AI Models for MRI-based Alzheimer's Disease Classification.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep learning, particularly convolutional neural networks (CNNs), is increasingly used for Alzheimer's disease (AD) classification and dementia severity assessment using 3D T1-weighted brain MRI.
• Existing research often relies on datasets like ADNI and NACC, primarily from North American and European populations, potentially limiting generalizability.

Purpose of the Study:

• To evaluate the interpretability of deep learning models for AD diagnosis by incorporating occlusion sensitivity analysis (OSA) and gradient-weighted class activation mapping (Grad-CAM).
• To assess the generalization capability of CNN models trained on North American/European datasets when applied to a diverse Indian cohort (NIMHANS).
• To investigate the impact of training CNN models on a combined dataset (ADNI/NACC + NIMHANS) for improved performance and generalizability.

Main Methods:

• Convolutional neural networks (CNNs) were employed for classifying Alzheimer's disease and inferring dementia severity from 3D T1-weighted brain MRI scans.
• Occlusion sensitivity analysis (OSA) and gradient-weighted class activation mapping (Grad-CAM) were integrated to enhance model interpretability.
• Model performance and generalizability were tested using datasets from North America/Europe (ADNI, NACC) and India (NIMHANS), including experiments with combined training data.

Main Results:

• Feature localization analyses using OSA and Grad-CAM demonstrated consistency with established knowledge of Alzheimer's disease pathology.
• The interpretability methods revealed distinct feature resolutions, aiding in the understanding of diagnostic inferences made by the CNN models.
• Experiments indicated that models trained on combined datasets may offer improved generalization to new populations.

Conclusions:

• OSA and Grad-CAM are valuable tools for interpreting deep learning models in Alzheimer's disease neuroimaging, providing insights into the features driving diagnostic predictions.
• Assessing model generalizability across diverse populations is crucial for the clinical translation of AI-based diagnostic tools.
• Combining diverse datasets for training CNNs shows promise for developing more robust and broadly applicable AI models for neurodegenerative disease diagnosis.