Differentiable neural architecture search for optimal spatial/temporal brain function network decomposition.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Decomposing brain's spatial/temporal function networks from 4D functional magnetic resonance imaging (fMRI) data is a key research topic.
• Deep neural networks offer advantages in fMRI data modeling but designing effective architectures for high-dimensional data remains challenging.

Purpose of the Study:

• Propose a novel spatial/temporal differentiable neural architecture search algorithm (ST-DARTS) for optimal brain network decomposition from fMRI data.
• Optimize the inner cell structure of recurrent neural networks (RNNs) for effective spatial/temporal brain network decomposition.

Main Methods:

• Developed ST-DARTS, a differentiable neural architecture search algorithm tailored for fMRI data.
• Optimized RNN cell structures to enhance spatial and temporal feature extraction.
• Introduced ST-DARTS+ with an early-stopping mechanism to improve efficiency.

Main Results:

• ST-DARTS demonstrated promising performance on seven fMRI tasks from the Human Connectome Project (HCP) dataset.
• Spatially, the model accurately identified stimuli-correlated brain network activation, similar to benchmarks.
• Temporally, the model's activity correlated highly with task designs.

Conclusions:

• ST-DARTS and ST-DARTS+ represent early advancements in using neural architecture search for optimal spatial/temporal brain network decomposition from fMRI.
• The proposed methods show significant promise for analyzing complex brain function networks.