Non-Rigid Cycle Consistent Bidirectional Network with Transformer for Unsupervised Deformable Functional Magnetic...

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Accurate inter-subject registration is crucial for statistical analysis in functional magnetic resonance imaging (fMRI) research.
• Traditional methods using structural MRI struggle with functional consistency as anatomical and functional regions may not align.
• Emerging functional-based methods often overlook valuable structural information.

Purpose of the Study:

• To develop an unsupervised, non-rigid registration method for fMRI that integrates both structural and functional information.
• To improve the accuracy and reliability of inter-subject fMRI registration for enhanced statistical analysis.
• To achieve superior structural and functional consistency in fMRI data.

Main Methods:

• Proposed a non-rigid, cycle-consistent, bidirectional network with Transformer architecture for unsupervised deformable fMRI registration.
• Integrated functional information by extracting local functional connectivity patterns and features.
• Employed a bidirectional network for forward and reverse registration and utilized Transformer for remote spatial mapping.

Main Results:

• The proposed method demonstrated improved registration performance compared to traditional (Affine, Syn) and learning-based (Transmorph-tiny, Cyclemorph, VoxelMorph x2) approaches.
• Achieved higher peak t-values and a greater number of suprathreshold voxels in key brain networks (DMN, VN, CEN, SMN) after registration.
• Showcased significant average improvements in peak t-value across multiple networks, indicating enhanced functional alignment.

Conclusions:

• The developed fMRI registration method effectively enhances structural and functional consistency between subjects.
• The integration of structural and functional information, facilitated by the Transformer-based network, leads to superior registration performance.
• This advancement holds promise for more robust statistical analyses in neuroscience research using fMRI data.