Hippocampal representations for deep learning on Alzheimer's disease

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Alzheimer's disease (AD) diagnosis can be improved by analyzing hippocampal changes.
• Deep learning models offer advanced analysis without manual feature engineering.
• The choice of input data representation for deep learning models in AD research is underexplored.

Purpose of the Study:

• To evaluate the impact of different hippocampal data representations on deep learning model performance for Alzheimer's disease prediction.
• To compare the interpretability of various representation-network pairs.
• To determine the optimal input format for predicting AD diagnosis and time-to-dementia.

Main Methods:

• Compared five distinct hippocampal representations (raw images, meshes, point clouds) with tailored deep learning architectures.
• Conducted rigorous evaluation for AD diagnosis and time-to-dementia prediction using an independent test dataset.
• Assessed the interpretability of each representation-network combination.

Main Results:

• The selection of hippocampal representation significantly influences the performance of deep learning models in predicting Alzheimer's disease.
• Different representations yielded varying degrees of diagnostic accuracy and prediction capabilities.
• Interpretability of the models was also found to be representation-dependent.

Conclusions:

• The choice of data representation is a critical factor in developing effective deep learning models for Alzheimer's disease analysis.
• Optimizing hippocampal representation can enhance both predictive accuracy and model interpretability for AD.
• Future research should carefully consider data input formats for deep learning applications in neurodegenerative disease.