RMT-match: an unsupervised 3D medical image registration network based on RMT and wavelet convolution

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Deformable image registration is vital for medical image analysis.
• Vision transformer (ViT) models for medical image registration (MIR) lack spatial priors in self-attention.
• Down-sampling operations in MIR can lead to loss of crucial spatial information, especially for 3D tasks.

Purpose of the Study:

• To propose a novel 3D deformable medical image registration (MIR) framework, RMT-Match.
• To enhance ViT-based MIR models by incorporating 3D spatial priors and improving down-sampling techniques.
• To balance model performance with computational efficiency in 3D MIR tasks.

Main Methods:

• Extended the Retentive Networks Meet vision transformers (RMT) structure to a 3D form (RMT-Match).
• Incorporated Manhattan distance into the spatial attenuation matrix to enhance self-attention with 3D spatial priors.
• Introduced a 3D wavelet convolutional down-sampling module for multi-frequency response, addressing information loss.

Main Results:

• RMT-Match demonstrated significant performance improvements over traditional CNN-based VoxelMorph on IXI and OASIS datasets (5.3% and 2.7% respectively).
• Outperformed state-of-the-art transformer-based TransMatch by 0.1% and 0.7% on tested datasets.
• Achieved a 40% reduction in parameter count compared to TransMatch, indicating improved computational efficiency.

Conclusions:

• The proposed RMT-Match framework effectively integrates 3D spatial priors and advanced down-sampling for deformable MIR.
• RMT-Match offers a promising approach for medical image registration, balancing high performance with computational efficiency.
• The method shows significant advantages and potential for various 3D medical image analysis applications.