Resting state network mapping in individuals using deep learning.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Resting state functional MRI (RS-fMRI) is crucial for clinical and research applications, including neurodegenerative disease analysis and pre-surgical planning.
• Reproducibility of RS-fMRI results often demands extensive, high-quality data, which is frequently unavailable.
• Localization of resting state networks (RSNs) is key for these applications.

Purpose of the Study:

• To develop and validate a novel deep 3D convolutional neural network (3DCNN) for voxelwise mapping of RSNs.
• To generate probabilistic RSN maps for individual participants using a large, publicly available dataset.
• To assess the robustness and clinical applicability of the developed 3DCNN method.

Main Methods:

• A deep 3D convolutional neural network (3DCNN) was trained on functional MRI data from 2010 healthy participants.
• Voxelwise mapping of RSNs was performed for each participant, generating probability maps.
• Mean and standard deviation probability maps were calculated and made publicly available. Results were compared to existing functional mapping techniques.

Main Results:

• The 3DCNN method successfully generated individual RSN localization maps.
• The approach demonstrated high resistance to noisy data and reduced RS-fMRI time points.
• Core regions within each RSN showed high average probability and low standard deviation, indicating reliability.

Conclusions:

• The 3DCNN algorithm provides a robust method for individual RSN localization, essential for clinical applications.
• The generated probabilistic maps and the algorithm's resilience to data limitations advance RS-fMRI utility.
• The alignment of 3DCNN RSN mapping with task-based fMRI findings supports functional network associations.