Abstract:
In laparoscopic liver surgery, overlaying preoperative information from a 3D computed tomography model with intraoperative reconstructed surfaces from laparoscopic images could support the surgeons by enhancing surgical navigation. However, achieving robust registration is difficult due to the lack of distinct features on the liver surface, the partial visibility of the liver in laparoscopic images, and the occurrence of disturbances like deforma- tions of the liver during surgery. Especially when attempting to register synthetic data with disturbances using handcrafted feature descriptors, identifying distinct features is particularly challenging. A promising solution is the utilization of learned feature descriptors. This work compares the registration accuracy of two learned feature descriptors, the PREDATOR and the LiverMatch, with currently used handcrafted feature descriptors. Additionally, the focus is on uncovering the reasons behind the superior performance of learned feature descriptors and assessing the degrees of disturbances for which the registration is still successful.To achieve these goals, data sets were generated for training, validation, and testing of the feature descriptors. While the 3D-IRCADb-01 data set was used for the preoperative data, the intraoperative data were synthetically generated. For the resulting features of the feature descriptors, the rigid transformation matrices were calculated with the random sample consensus algorithm, and the registration accuracy was evaluated with the root mean square error (RMSE). Additionally, the features were analyzed with regard to the possibility of calculating correspondences. To further understand the feature calculation of the learned feature descriptors, the features of the preoperative data were analyzed.The examined learned feature descriptors significantly improved the registration accuracy. One investigated reason was the ability of learned feature descriptors to identify more corre- spondences that aligned with the actual correspondences. Similar to the handcrafted feature descriptors, the learned ones used geometric properties for the feature calculation. Addition- ally, they benefit from the exchange of information between preoperative and intraoperative data, which enabled them to compute more distinctive features. When assessing the impact of disturbances individually, the LiverMatch achieved RMSE values of 5.38 mm for large variations. Regarding combinations of all disturbances, the LiverMatch reached its limits more quickly, with the patch size notably affecting the registration accuracy.In summary, this study shows the large potential of the examined learned feature descrip- tors for use in laparoscopic liver surgery because they enabled a more precise initial rigid alignment compared to their handcrafted counterparts. Furthermore, by optimizing both the training data set and hyperparameters, there is the potential for further improving the results.
