An interpretable multimodal deep learning framework for Alzheimer's disease diagnosis

Area of Science:

• Artificial Intelligence in Medical Imaging
• Neuroscience and Neuroimaging
• Biomedical Data Science

Background:

• Alzheimer's disease (AD) poses a growing public health challenge, with early neurodegeneration often missed by current diagnostics.
• Medical imaging, including MRI and PET, offers insights into AD-related brain changes but integration into AI frameworks is limited.
• Developing explainable AI for multimodal imaging is crucial for advancing AD diagnosis.

Purpose of the Study:

• To develop and evaluate NeuroFusion-ADNet, a novel AI model for improved Alzheimer's disease diagnosis.
• To enhance diagnostic accuracy and clinical interpretability by combining structural MRI and functional PET data.
• To create a transparent and explainable AI framework for early AD detection.

Main Methods:

• A dual-path deep learning model, NeuroFusion-ADNet, was designed to jointly process MRI and PET data.
• The model incorporates modality-specific encoders, a cross-attention fusion layer, and a segmentation-informed classification module.
• Explainability features, including attention heatmaps and Local Interpretable Model-Agnostic Explanations, were integrated.

Main Results:

• NeuroFusion-ADNet achieved 99.48% classification accuracy and a 0.985 Dice coefficient, outperforming baseline models.
• Attention visualizations highlighted the model's focus on clinically relevant brain regions like the hippocampus and entorhinal cortex.
• Ablation studies confirmed the significance of individual architectural components.

Conclusions:

• This study presents a clinically promising, multimodal AI framework for Alzheimer's disease diagnosis.
• NeuroFusion-ADNet enhances diagnostic accuracy and transparency through explainable AI techniques.
• The developed framework provides a foundation for efficient, interpretable, and deployable tools for early AD detection.