CorrMorph: Unsupervised Deformable Brain MRI Registration Based on Correlation Mining.

Area of Science:

• Medical Image Analysis
• Neuroimaging
• Computer Vision

Background:

• Deformable image registration is crucial for medical image analysis.
• Existing methods struggle with spatial misalignment, style discrepancies, and capturing inter-image correlations.
• CNNs excel at local features but miss global context, while Transformers capture global context but miss local details.

Purpose of the Study:

• To develop an unsupervised deformable brain MRI registration network, CorrMorph.
• To address limitations of single-stream and dual-stream networks, and CNN/Transformer-based approaches.
• To improve accuracy and robustness in brain MRI registration.

Main Methods:

• Proposed CorrMorph, an unsupervised deformable brain MRI registration network.
• Introduced a match-fusion strategy for independent shallow feature extraction and deeper correlation learning.
• Developed Correlation Matching Module (CMM) for feature matching and Feature Transmission Module (FTM) for spatial feature extraction.

Main Results:

• CorrMorph achieved state-of-the-art performance on three brain MRI datasets.
• Demonstrated an average improvement of 2.7% in Dice Similarity Coefficient (DSC) over VoxelMorph.
• Effectively addressed challenges of spatial misalignment, style discrepancy, and complex spatial correspondences.

Conclusions:

• CorrMorph offers a robust and accurate solution for unsupervised deformable brain MRI registration.
• The proposed match-fusion strategy, CMM, and FTM modules significantly enhance registration performance.
• This work advances the field of medical image registration, particularly for brain MRI analysis.