Functional magnetic resonance imaging progressive deformable registration based on a cascaded convolutional neural...

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• Accurate intersubject registration of functional magnetic resonance imaging (fMRI) is crucial for robust group analysis.
• Traditional registration methods using structural images or manual feature extraction have limitations in achieving functional alignment.
• Deep learning approaches show promise for deformable image registration in neuroimaging.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for precise deformable image registration of fMRI data.
• To improve functional alignment across subjects for enhanced statistical analysis in neuroimaging studies.

Main Methods:

• A three-cascaded multi-resolution network (MR-Net) was proposed for deformable image registration.
• MR-Net employs a two-stream architecture to extract features from moving and fixed images separately.
• The framework predicts cascaded sub-deformation fields with smoothness constraints to ensure topological correctness.

Main Results:

• The MR-Net was validated on the 1000 Functional Connectomes Project (FCP) and Eyes Open Eyes Closed fMRI datasets.
• The method significantly increased peak t values in six brain functional networks.
• Compared to traditional methods (FSL, SPM) and deep learning networks (VM, VTN), MR-Net demonstrated superior performance, with improvements ranging from 11.88% to 47.58%.

Conclusions:

• The proposed three-cascaded MR-Net effectively achieves statistically significant improvements in functional consistency across subjects.
• This deep learning framework offers a more accurate and efficient solution for fMRI intersubject registration.