Hierarchical cumulative network for unsupervised medical image registration

Area of Science:

• Medical image analysis
• Deep learning in medical imaging
• Computational anatomy

Background:

• Deformable medical image registration is crucial for analyzing anatomical changes.
• Existing unsupervised deep learning methods often fail to find optimal similarity positions.
• This limitation can lead to suboptimal registration accuracy.

Purpose of the Study:

• To introduce a novel Hierarchical Cumulative Network (HCN) for unsupervised deformable medical image registration.
• To address the limitation of overlooking optimal similarity positions in current methods.
• To improve the accuracy and robustness of medical image alignment.

Main Methods:

• Developed a novel Hierarchical Cumulative Network (HCN) incorporating a Bidirectional Asymmetric Registration Module (BARM).
• BARM learns asymmetric displacement vector fields (DVFs) to optimally warp both moving and fixed images.
• Integrated BARM into a Laplacian pyramid network for hierarchical, recursive warping and DVF accumulation.

Main Results:

• The proposed HCN significantly outperformed traditional and state-of-the-art registration methods on 3D Brain MRI datasets.
• HCN demonstrated strong performance on the MICCAI Learn2Reg 2021 challenge validation set.
• Cross-dataset evaluation confirmed the excellent generalization capabilities of the HCN method.

Conclusions:

• The HCN is an effective unsupervised deep learning approach for deformable medical image registration.
• The Bidirectional Asymmetric Registration Module (BARM) successfully addresses the optimal similarity position challenge.
• HCN achieves high accuracy and excellent generalization, making it suitable for diverse clinical applications.