Learning 3D Feature Descriptors for the Registration in Laparoscopic Liver Surgery
The registration of a preoperative 3D model with the 3D reconstructed intraoperative laparoscopic image in laparoscopic liver surgery is an unsolved problem. Occurring disturbances, e.g. deformation due to the intraoperative gas pressure and spatial displacement due to the use of different imaging modalities, complicate the alignment of both data. One important step of the registration pipeline is the feature description, in which diverse geometrical information of the points clouds are encoded. In the past, classic 3D feature descriptors were compared with each other. Those methods use predefined information as the point density or depth information from one or several point of views. In recent years, besides classic approaches, learning-based feature descriptors were introduced. Those descriptors learn which information are suitable to encode the point cloud structure unambiguously. Thus, we hypothesis that the use of a learning 3D feature descriptor in laparoscopic liver surgery can improve the registration accuracy. The aim of this work is to train, test and validate a promising learning 3D feature descriptor. Thereby, the training data consists of simulated pseudo-intraoperative surface patches and the corresponding preoperative 3D models. Afterwards, the learning feature descriptor should be used to register both point sets. The registration accuracy is evaluated and compared with the accuracy of the classic registration pipeline.