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

Pyramid-based deformable registration networks achieve strong accuracy by decomposing complex deformations into a stack of subfields. However, most methods only rely on a single-step prediction at each scale and lack explicit modeling of intra-level motion ambiguity. Additionally, Transformer-based methods typically require millions of parameters. Such limitations restrict their ability to deploy, and jointly handle large displacements and subtle local motions. We propose ESR-Net, a parameter-efficient and motion-aware registration network that introduces an Exploratory Self-Refinement Module (ESRM) at each decoder level. A four-stage design in ESRM explicitly captures, guides, evaluates, and refines diverse motion possibilities rather than relying on a direct deformation estimation. In particular, ESRM first predicts motion-aware offset fields via spatial-channel attention, then dynamically relocates and updates warped moving features via deformable convolution, estimates confidence-aware weights by comparing updated moving and fixed features, and finally performs confidence-weighted fusion of multiple offset candidates to produce refined contextual representations and deformation subfields. This design enables ESR-Net to jointly capture intra-level motion diversity and inter-level dependencies, allowing it to handle both large displacements and subtle local deformations with high precision in a purely convolutional and lightweight manner. Extensive experiments on three 3D brain MRI datasets and one lung CT dataset show that ESR-Net outperforms popular CNN-based, Transformer-based, and pyramid-based methods while requiring only 0.60M parameters. These results demonstrate that explicit exploratory self-refinement offers an efficient and effective alternative to heavy Transformer-based registration models. Our code is publicly available at https://github.com/YoboY23420/ESR-Net.