A markerless registration method for neurosurgical navigation: a phantom study

Purpose Landmark registration is the traditional method to convert patient's physical space to image space for neurosurgical navigation. While marker-based registration is prevailing, its shortcomings such as marker shift and missing often cause registration inaccuracy or failure. Now, commercially available markerless registration which uses laser surface scanner or touch sensitive point scanner became popular [1], but an extra dynamic reference base (DRB) is needed to be attached with the scanner for calibration, which results in complexity of registration procedure. A markerless registration method without dedicated scanner is proposed for neurosurgical navigation. Methods One head phantom with seven fiducials China) attached on the surface as targets was used in the experiment. CT images were acquired by one Philips Brilliance CT 64-scanner with 0.7x0.7x0.5mm resolution, then skin surface and 3D sagittal-curve were extracted by thresholding method. In order to get the 3D sagittal-curve, the centroid and the cutting plane in sagittal orientation were computed, then the points from the contours of this plane whose z coordinate larger than that of nasal tip (the corresponding pixel value in y coordinate is minimal) were extracted. Mesh was simplified based on error quadrics and directed edges [2]. One Polaris Spectra hybrid optical tracking system (Northern Digital Inc., Waterloo, ON, Canada) and a common tracked probe were used to acquire intra-operative position of the head phantom. The registration procedure is accomplished based on matching the intra-operative swabbed lines with pre-operative skin surface (see Fig. 1). Firstly, Gaussian mixture model (GMM) [3] was integrated for matching pre-operative 3D sagittal-curve with the simplified homologous point sets acquired by the probe, the points too close to their neighbors were abandoned. Then another acquired intra-operative swabbed point sets were aligned with simplified skin surface to refine the coarse registration through perturbed iterative closest point (ICP) [4] to prevent local minimum problem. Results Two experiments were conducted: one with marker based registration approach, the other with markerless based registration method. The target registration error (TRE)s of 7 fiducials were calculated for seven repeated experiments, their mean and standard deviation are shown in Fig. 2, the average TRE are 1.77±0.78mm and 1.94±0.90mm for landmark and markerless registration respectively. Paired t-test for 49 targets shows that there is no significant difference in TRE between marker and markerless registration method (t=1.15, p=0.26>0.05). The markerless registration took about 30s on one common personal computer. Conclusion A novel markerless registration method for neurosurgical navigation was developed …