MeiBRD uses meta-learning to correct liver deformation during surgery

Accurate soft-tissue registration during surgery is notoriously difficult due to large deformations and limited intraoperative data. Biomechanical models provide physical priors but suffer from persistent prediction bias due to simplifying assumptions, while purely data-driven methods struggle with data efficiency and generalization. MeiBRD offers a hybrid solution: instead of learning the full deformation field, it learns a residual deformation function that corrects linear biomechanical predictions. The residual is modeled as a graph neural diffusion function with geometry-aware attention over the 3D liver mesh, enabling information to propagate across the organ surface.

A key innovation is treating sparse intraoperative measurements as 'context samples' where input-output pairs of the residual function are fully observed. This reformulates the problem as a meta-learning task: the model learns to learn the residual function from these limited intraoperative samples using feedforward meta-learners. Experiments on a deformable liver phantom dataset show that MeiBRD significantly outperforms rigid, biomechanical, and data-driven baselines in registration accuracy and generalization, particularly when faced with out-of-distribution geometries and deformations. This hybrid approach promises more reliable surgical guidance with minimal data requirements.