engleski

Improving Markerless Registration Accuracy by Mapping Facial Deformation

Robotics is gaining an increasing role in stereotactic neurosurgery. Novel methods for intraoperative localization of the patient should be developed for tackling the challenges of the localization and registration processes simplification. Most state of the art robotic systems use marker-based localization and registration while markerless localization is at its infancy. One of the challenges of markerless localization and registration is face deformation. The human skin and parts of the human face can deform from the preoperative scanning phase to the operative position in the actual surgery due to swelling, different orientations of scanning and operation, etc. The human face is not uniformly deformable, with some parts of the face deforming less than others for example due to the differences in skin thickness, and the depth of the bone with respect to the outer surface of the skin. This presents an opportunity for improving the markerless registration by identifying and using less deformable parts of the face while not registering parts of the face that are more deformable. In this paper, an algorithm for dividing the face surface into multiple regions and determining each region's deformability is proposed and validated on preliminary data from actual neurosurgical operations. Two approaches for determining the facial regions are proposed. The first approach creates a uniform n×n 3D grid and superimposes it over the point cloud of the patient’s face. The second approach utilizes the detection of facial landmarks in a 2D image and maps these landmarks to the point cloud, where each of the landmarks corresponds to one distinct facial region. To calculate the deformability of any facial region, multiple point clouds can be registered using ground truth measurements, followed by their respective deformability calculation as a standard deviation of all point clouds. Ground truth registration is calculated using bone attached fiducial markers. A statistical measure of the registration of different regions is used to validate the best regions to be used in each approach and to validate each approach. Preliminary results are given for one intraoperative point cloud. Error is measured as the average Euclidean distance from registered cloud markers to ground truth markers. When the whole face was used for registration, the error was 4.19mm, while when using only selected parts of the face, the error was reduced to 1.78mm.

registration ; point cloud ; deformation

nije evidentirano