Preoperative and intraoperative laparoscopic liver surface registration using deep graph matching of representative overlapping points.

Purpose: In laparoscopic liver surgery, registering preoperative CT-extracted 3D models with intraoperative laparoscopic video reconstructions of the liver surface can help surgeons predict critical liver anatomy. However, the registration process is challenged by non-rigid deformation of the organ due to intraoperative pneumoperitoneum pressure, partial visibility of the liver surface, and surface reconstruction noise.

Methods: First, we learn point-by-point descriptors and encode location information to alleviate the limitations of descriptors in location perception. In addition, we introduce a GeoTransformer to enhance the geometry perception to cope with the problem of inconspicuous liver surface features. Finally, we construct a deep graph matching module to optimize the descriptors and learn overlap masks to robustly estimate the transformation parameters based on representative overlap points.

Results: Evaluation of our method with comparative methods on both simulated and real datasets shows that our method achieves state-of-the-art results, realizing the lowest surface registration error(SRE) 4.12 mm with the highest inlier ratios (IR) 53.31% and match scores (MS) 28.17%.

Conclusion: Highly accurate and robust initialized registration obtained from partial information can be achieved while meeting the speed requirement. Non-rigid registration can further enhance the accuracy of the registration process on this basis.