Automatic recognition of holistic functional brain networks using iteratively optimized convolutional neural networks...

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• fMRI data decomposition advanced from shallow models (ICA, SCDL) to deep learning (DBN, DCAE).
• Interpreting decomposed networks is challenging due to lack of atlases, cross-subject correspondence issues, and individual variability.
• Deep convolutional neural networks (CNNs) show promise for fMRI network classification but require extensive, accurate manual labeling for training data.

Purpose of the Study:

• To develop an automated framework for functional brain network recognition and training data labeling.
• To address the infeasibility of manual labeling for large-scale fMRI datasets.
• To enable fully automatic, large-scale classification of functional brain networks.

Main Methods:

• Proposed a novel iteratively optimized deep learning CNN (IO-CNN) framework.
• Implemented an automatic weak label initialization strategy.
• Utilized a large-scale fMRI dataset (ABIDE-II, 1099 brains) for experiments.

Main Results:

• The IO-CNN framework successfully automated functional brain network recognition and labeling.
• Achieved efficient, large-scale classification of functional brain networks.
• Demonstrated the framework's promise on a substantial fMRI dataset.

Conclusions:

• The proposed IO-CNN framework effectively automates functional brain network recognition.
• This approach overcomes significant hurdles in training data preparation for deep learning models in neuroimaging.
• The framework holds great promise for advancing large-scale analysis of functional brain networks.