Data-efficient resting-state functional magnetic resonance imaging brain mapping with deep learning.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Brain Network Analysis

Background:

• Resting-state functional MRI (RS-fMRI) is crucial for presurgical mapping of eloquent cortex.
• Current RS-fMRI mapping requires significant scanning time, impacting patient comfort and costs.
• Developing faster, accurate methods for mapping resting-state networks (RSNs) is essential.

Purpose of the Study:

• To develop a deep 3D convolutional neural network (3DCNN) for voxel-wise mapping of language (LAN) and motor (MOT) RSNs.
• To achieve reliable RSN mapping using minimal RS-fMRI data.
• To enhance presurgical planning efficiency and accuracy.

Main Methods:

• A 3DCNN was trained using RS-fMRI data from 2252 healthy adults and publicly available datasets.
• Model performance was validated using varying amounts of RS-fMRI data and tested on patients with glioblastoma multiforme.
• Random permutations of RS-fMRI regions of interest were employed for model training.

Main Results:

• The 3DCNN achieved 96% out-of-sample validation accuracy across diverse data.
• Accurate LAN and MOT RSN mapping was achieved with as little as 2.5 minutes of RS-fMRI data (96.9% LAN, 96.3% MOT true-positive rate).
• The model successfully mapped RSNs in brain tumor patients despite structural and functional alterations.

Conclusions:

• The developed 3DCNN provides a highly efficient method for presurgical functional mapping.
• Functional maps generated by the 3DCNN can significantly inform surgical planning for brain tumor patients.
• This approach promises improved functional preservation in neurosurgical oncology.