Constructing fine-granularity functional brain network atlases via deep convolutional autoencoder.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Functional brain network reconstruction from fMRI data is challenging due to individual variability and scale.
• Existing methods like ICA and sparse coding struggle to derive group-wise common network atlases.
• Deep convolutional neural networks (CNNs) offer superior spatial pattern description capabilities.

Purpose of the Study:

• To develop a novel deep 3D convolutional autoencoder (CAE) network for effective extraction of spatial brain network features.
• To create an Apache Spark-enabled framework for fast clustering of large numbers of network maps into fine-granularity atlases.
• To refine existing brain atlases and identify subtle network abnormalities.

Main Methods:

• A deep 3D CAE network was designed to extract spatial features from functional brain networks.
• An Apache Spark framework was developed for efficient clustering of network maps.
• Trained CAE models on Human Connectome Project (HCP) fMRI data, refining 10 resting state networks (RSNs) into 17 fine-granularity atlases.
• Applied the method to a mild traumatic brain injury (mTBI) dataset.

Main Results:

• The deep CAE framework successfully refined 10 RSNs into 17 fine-granularity atlases.
• Learned features corrected mislabeled outliers in training data.
• Identified unique network patterns specific to different brain task states.
• Effectively detected abnormal small networks in mTBI patients compared to controls.

Conclusions:

• The proposed deep learning and big data analysis solution offers a promising approach for modeling functional connectomes with fine granularity using fMRI data.
• The framework demonstrates potential for atlas refinement, outlier detection, and identifying subtle network alterations in neurological conditions.
• This method advances the creation of detailed brain network atlases for research and clinical applications.