TransMatch: A Transformer-Based Multilevel Dual-Stream Feature Matching Network for Unsupervised Deformable Image...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Feature matching is essential for image registration, but traditional methods are slow and application-specific.
• Learning-based methods like VoxelMorph and TransMorph show promise but use implicit feature matching.
• Existing single-stream deep learning models concatenate images, limiting explicit inter-image feature relationship understanding.

Purpose of the Study:

• To introduce TransMatch, a novel end-to-end dual-stream unsupervised framework for deformable medical image registration.
• To enable explicit multilevel feature matching between image pairs using Transformer self-attention.
• To achieve state-of-the-art performance in deformable image registration.

Main Methods:

• Proposed a dual-stream architecture where each input image is processed independently.
• Implemented explicit multilevel feature matching using the query-key mechanism from Transformer self-attention.
• Conducted experiments on three 3D brain MRI datasets: LPBA40, IXI, and OASIS.

Main Results:

• TransMatch achieved state-of-the-art performance across multiple evaluation metrics.
• The method demonstrated superior results compared to established registration techniques (SyN, NiftyReg, VoxelMorph, etc.).
• The dual-stream approach with explicit feature matching proved effective for 3D brain MR image registration.

Conclusions:

• The proposed TransMatch framework offers a significant advancement in unsupervised deformable medical image registration.
• Explicit feature matching via Transformer self-attention enhances registration accuracy and robustness.
• TransMatch provides a competitive and effective alternative to existing registration methodologies.