Identity-mapping cascaded network for fMRI registration

Area of Science:

• Neuroscience
• Medical Imaging
• Computer Vision

Background:

• Accurate inter-subject registration of functional magnetic resonance imaging (fMRI) is crucial for enhancing statistical power in group analyses.
• Deep learning methods show promise for improving the accuracy and efficiency of fMRI image registration.

Purpose of the Study:

• To propose and evaluate a novel deep learning-based registration network, the 30-Identity-Mapping Cascaded network (30-IMCNet), for resting-state fMRI (rs-fMRI) and task-related fMRI.
• To demonstrate the superiority of 30-IMCNet compared to traditional and existing deep learning registration methods.

Main Methods:

• Developed a 30-Identity-Mapping Cascaded network (30-IMCNet) featuring a cascaded architecture with identity-mapping paths for progressive image warping.
• Implemented and tested 30-IMCNet on the 1000 Functional Connectomes Project (rs-fMRI) and Eyes Open Eyes Closed (task-fMRI) datasets.
• Evaluated registration quality using group-level analysis metrics including peak t-value, cluster-level evaluation, intersubject functional network correlation, ALFF, and ReHo.

Main Results:

• 30-IMCNet demonstrated significant improvements in peak t-value compared to FSL, SPM, and other deep learning methods, with gains of 48.90%, 30.73%, 36.38%, and 16.73% respectively.
• The method achieved superior functional registration performance, leading to enhanced functional consistency in group analyses.
• Evaluations on both rs-fMRI and task-fMRI datasets confirmed the effectiveness of the proposed network.

Conclusions:

• The 30-Identity-Mapping Cascaded network (30-IMCNet) offers a robust and effective solution for inter-subject fMRI registration.
• This advancement in registration accuracy can significantly improve the reliability and statistical power of neuroimaging group studies.
• The proposed method represents a notable step forward in applying deep learning to neuroimaging data analysis.