A Parameter-efficient and Motion-aware Exploratory Self-Refinement Network for 3D Brain MRI Registration.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• Pyramid-based deformable registration networks offer accuracy by decomposing deformations.
• Existing methods often lack explicit modeling of intra-level motion ambiguity and can be parameter-heavy.
• Transformer-based methods, while powerful, typically require millions of parameters, limiting deployment.

Purpose of the Study:

• To propose ESR-Net, a parameter-efficient and motion-aware registration network.
• To introduce an Exploratory Self-Refinement Module (ESRM) for improved deformation estimation.
• To enable joint handling of large displacements and subtle local motions in medical image registration.

Main Methods:

• ESR-Net utilizes a four-stage Exploratory Self-Refinement Module (ESRM) at each decoder level.
• ESRM explicitly captures, guides, evaluates, and refines diverse motion possibilities.
• Methods include spatial-channel attention, deformable convolution, confidence-aware weighting, and confidence-weighted fusion.

Main Results:

• ESR-Net demonstrates superior performance on 3D brain MRI and lung CT datasets.
• The network outperforms popular CNN-based, Transformer-based, and pyramid-based registration methods.
• ESR-Net achieves high precision in handling both large displacements and subtle local deformations with only 0.60M parameters.

Conclusions:

• Explicit exploratory self-refinement provides an efficient and effective alternative to heavy Transformer-based registration models.
• ESR-Net offers a lightweight yet powerful solution for medical image registration.
• The proposed method enables precise registration with significantly reduced parameter count.