On Studying Transformer Networks for Volume-to-Surface Registration of Inhomogeneous Soft Bodies for Liver Laparoscopy

An important practical consideration in laparoscopic liver surgery is the limited visual information relative to open surgery. In laparoscopic interventions, the surgeon’s view of the liver surface is generally limited to the scene provided by a single scope with a narrow field-of-view. This limits the ability to navigate towards internal lesions previously identified through pre-procedural imaging. Surgical navigation during laparoscopy could be enhanced by registration of full preoperative liver models derived from pre-procedural imaging scans onto the partial laparoscopic view of the liver. This entails both a rigid registration to match the liver surface view to the preoperative volume, as well as a nonrigid registration to correct for the deformation between the pre- and intraoperative scenarios. Prior work has demonstrated the feasibility of data-driven methods for both tasks. However, both registrations are impeded by intraoperative occlusion, limited surface landmarks, and movement of the liver due to abdomen insufflation and patient breathing. In this work, we extend state-of-the-art deep learning frameworks for the task of nonrigid registration. In particular, we investigate the use of vision transformer attention blocks within a U-Net structure to improve the prediction of a displacement field between the preand intraoperative surfaces. We also investigate the robustness of these networks by creating novel training data with imperfect rigid registration and inhomogeneous mechanical properties. Our results indicate that the addition of rigidly deformed training data improves network performance regardless of rigid transformation in the test set. Specifically, we show that a network trained with rigidly transformed data can achieve displacement prediction errors of less than 5 mm on a simulated liver task. Contrary to expectations, the use of transformer architectures and training on inhomogeneous data each reduce network performance across nearly all cases. This work highlights the advantages of a neural network registration method: it requires no knowledge of boundary conditions, has no reliance on manually-tuned parameters, and demonstrates robustness towards sub-optimal prior rigid registration. This work expands the corpus of research backing data-driven volume-to-surface registration as a potentially powerful tool in the advancement of surgical navigation.