An explainable deep learning framework for characterizing and interpreting human brain states.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Image Analysis

Background:

• Deep learning models in medical imaging often lack interpretability.
• Existing brain imaging studies using deep learning struggle with model explainability.

Purpose of the Study:

• To develop a novel, interpretable deep learning model for human brain state characterization and prediction.
• To integrate domain knowledge from structural and functional brain connectivity into a graph convolutional neural network.

Main Methods:

• A self-attention graph pooling (SAGPool)-based graph convolutional neural network was developed.
• The model integrated the dense individualized and common connectivity-based cortical landmarks system (DICCCOL) and holistic atlases of functional networks and interactions system (HAFNI).
• Experiments were conducted on the Human Connectome Project (HCP) dataset.

Main Results:

• The model achieved high classification performance, with 93.7% accuracy for seven-task classification and 100% for binary classification.
• The importance of brain regions contributing to classification was quantified and visualized.
• Neuroscientific interpretation confirmed the explainability of the model.

Conclusions:

• The proposed domain knowledge-informed SAGPool model offers a promising approach for interpretable brain state analysis.
• The model successfully characterizes and interprets human brain states by leveraging structural and functional connectivity data.
• The explainability of the model aids in understanding the neuroscientific basis of brain states.