Improving Intraoperative Liver Registration in Image-Guided Surgery with Learning-Based Reconstruction

Nonrigid image-to-physical registration is a crucial component in image-guided liver surgery. To overcome the problems caused by noisy, partial, and sparse intraoperative sampling, we propose a novel occupancy-learning-based mesh to point cloud registration and apply it to align the preoperative liver image to intraoperative samples. We train a point cloud deep network to reconstruct occupancy function from sparse points and use this reconstructed liver to guide the nonrigid registration. Experiments show this method reduces Target Registration Error (TRE) of rigid and nonrigid baselines by 21.5% and 11.8%. For training this occupancy network, a novel crossover method is proposed to synthesize deformed liver meshes and points. We demonstrate that the system is robust to the imperfectness of generated training data. This method might be useful in other areas that require soft-tissue registration, where only very sparse data is available during acquisition.