Modeling task-based fMRI data via deep belief network with neural architecture search.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Deep neural networks (DNNs) offer superior representation for fMRI data compared to traditional methods.
• Manual design of DNN architectures for high-dimensional fMRI data is time-consuming and suboptimal.
• Developing automated methods for fMRI data analysis is crucial for advancing neuroscience research.

Purpose of the Study:

• To propose an unsupervised neural architecture search (NAS) framework, named NAS-DBN, for modeling volumetric fMRI data.
• To automate the discovery of optimal Deep Belief Network (DBN) architectures for fMRI analysis.
• To improve the identification and characterization of functional brain networks (FBNs) and their temporal dynamics.

Main Methods:

• Developed NAS-DBN, an unsupervised NAS framework utilizing Particle Swarm Optimization (PSO) for DBN architecture search.
• Applied the NAS-DBN framework to model volumetric task fMRI data.
• Compared NAS-DBN performance against manually designed DBNs and established methods like GLM and ICA.

Main Results:

• NAS-DBN rapidly identified a robust DBN architecture, outperforming manual designs with a 47.9% improvement in performance (testing loss of 0.0197).
• Identified 260 FBNs, including task-specific networks and resting-state networks (RSNs), with high overlap rates to GLM and ICA templates (up to 63.9% improvement).
• Demonstrated hierarchical organization of FBNs and accurate temporal response generation that closely matched task designs.

Conclusions:

• The NAS-DBN framework provides an effective and unsupervised approach for optimizing neural network architectures in volumetric fMRI data analysis.
• This method significantly enhances the modeling of functional brain networks and their temporal characteristics.
• NAS-DBN offers a promising direction for automated and efficient analysis of complex neuroimaging data.