MeiBRD: Meta-Learning Intraoperative Biomechanical Residual Deformation

The article presents MeiBRD, a hybrid registration framework designed to enhance intraoperative liver registration by addressing soft-tissue deformation and sparse measurements. It employs a graph neural network with a geometry-aware attention mechanism to model a residual deformation function that corrects linear biomechanical predictions, utilizing sparse intraoperative correspondences as context samples for training. Experimental results indicate that MeiBRD significantly improves registration accuracy and generalization, particularly for challenging geometries and deformations, which is crucial for practitioners aiming to enhance surgical navigation and planning in dynamic environments.