Alzheimer's disease prediction via an explainable CNN using genetic algorithm and SHAP values.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Neuroscience

Background:

• Deep learning models like 3D-CNNs excel at image classification but lack transparency.
• Interpreting black-box models is challenging in healthcare, particularly for diagnosing diseases like Alzheimer's.
• Explainable AI (XAI) techniques aim to enhance model interpretability.

Purpose of the Study:

• To introduce GASHAP, a novel XAI technique integrating genetic algorithms (GA) with SHAP, to improve 3D-CNN explainability.
• To enhance the transparency of 3D-CNN models used for classifying Alzheimer's disease in MRI scans.
• To provide diagnostic insights at the level of anatomically defined brain regions, moving beyond voxel-level analysis.

Main Methods:

• Implemented a 3D-CNN for classifying Alzheimer's disease in MRI brain scans.
• Applied the GASHAP technique, combining GA and SHAP, to identify significant brain regions.
• Generated a definitive brain mask highlighting critical regions for Alzheimer's diagnosis.

Main Results:

• Developed a 3D-CNN model for Alzheimer's disease classification from MRI scans.
• Successfully applied GASHAP to enhance model transparency and interpretability.
• Produced a brain mask pinpointing key anatomical regions for Alzheimer's diagnosis.

Conclusions:

• GASHAP offers improved explainability for 3D-CNNs in medical imaging.
• The technique aids in identifying crucial brain regions for Alzheimer's disease diagnosis.
• This approach enhances the clinical utility of AI in neurodegenerative disease detection.