Physics-Visual Fusion for Cross-Posture Deformation Mapping in Robotic Breast Ultrasound

Robotic breast ultrasound scanning aids early cancer screening but faces significant challenges from substantial soft tissue deformations induced by supine-to-lateral postural changes. These deformations cause deviations in pre-planned scan paths and lesion drifts. Consequently, simple re-planning becomes infeasible due to the prohibitive time cost of global re-scanning and the risk of lesion identity ambiguity arising from spatial correspondence loss. To address this, we propose a physics-visual fusion framework for accurate cross-posture deformation mapping. First, a geometry-specific 3D breast mechanical baseline is reconstructed in the supine posture using multi-view RGB-D perception, eliminating the need for medical imaging priors. Next, phantom-specific model-equivalent elastic parameters are fitted via Bayesian Optimization by minimizing the geometric discrepancy between FEM simulations and single-view surface point clouds across multiple lateral postures. Finally, a continuous deformation mapping is achieved by coupling the FEM-predicted, gravity-driven deformations with a spring-type registration force field, which applies local non-rigid corrections based on target-posture point cloud observations. Controlled single-phantom experiments show that our method balances surface fitting and spatial correspondence preservation, improves surface alignment over pure FEM, and reduces the mean scan-path re-mapping and lesion localization errors by 60.84% and 62.81% compared with rigid registration and by 79.14% and 79.38% compared with CPD, respectively.