Frequency-Aware Interpretable Deep Learning Framework for Alzheimer's Disease Classification Using rs-fMRI.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Alzheimer's disease (AD) diagnosis relies on identifying alterations in brain connectivity.
• Understanding spectral and temporal changes in brain networks is crucial for developing effective biomarkers.

Purpose of the Study:

• To introduce FINE (Frequency-aware Interpretable Neural Encoder), a novel deep learning model.
• To capture multi-scale temporal and frequency-specific patterns in dynamic functional network connectivity (dFNC) from resting-state fMRI data.
• To enhance Alzheimer's disease classification and provide interpretable insights into brain connectivity disruptions.

Main Methods:

• Developed FINE, a deep learning model integrating convolutional layers, wavelet layers, transformers, and static encoders.
• Employed an end-to-end framework for joint modeling of temporal evolution and spectral content of brain networks.
• Utilized gradient-based saliency maps for frequency-wise interpretability, aligned with statistical group differences.

Main Results:

• Evaluated on the OASIS-3 dataset (856 subjects), FINE achieved an ROC-AUC of 0.769 for AD classification.
• Identified frequency-specific connectivity disruptions in subcortical, sensorimotor, and cerebellar networks.
• Demonstrated the model's ability to reveal potential robust, biologically meaningful biomarkers for AD.

Conclusions:

• Frequency-aware modeling and interpretable architectures can significantly advance AD classification.
• FINE provides valuable insights into the functional disruption of AD-related brain dynamics.
• The approach holds promise for developing more informative biomarkers and deepening the understanding of Alzheimer's disease mechanisms.