Mamba-based deformable medical image registration with an annotated brain MR-CT dataset

Area of Science:

• Medical Image Analysis
• Neuroimaging
• Artificial Intelligence in Medicine

Background:

• Deformable registration is crucial for multi-modal neuroimaging analysis.
• Current learning-based methods face challenges in accuracy and efficiency for brain MR-CT registration.
• A need exists for improved benchmarks and methods for brain MR-CT registration.

Purpose of the Study:

• To introduce SR-Reg, a novel benchmark dataset for brain MR-CT registration.
• To present MambaMorph, a new deformable registration network utilizing a state space model.
• To evaluate MambaMorph's performance against existing advanced networks.

Main Methods:

• Developed the SR-Reg dataset with 180 paired volumetric MR-CT images and annotations.
• Proposed MambaMorph, a network integrating Mamba for global feature learning and a fine-grained extractor.
• Conducted experiments on multi- and mono-modal registration tasks.

Main Results:

• MambaMorph demonstrated superior performance compared to ConvNet-based and Transformer-based networks.
• The proposed method achieved significant improvements in both accuracy and efficiency.
• SR-Reg dataset facilitates further research in brain MR-CT registration.

Conclusions:

• MambaMorph offers a promising advancement in deformable image registration for neuroimaging.
• The SR-Reg dataset and MambaMorph contribute to the field of multi-modal medical image analysis.
• The study highlights the potential of state space models in medical image registration.