A phantom study for the validation of a Surgical Navigation System Based on Real-Time Segmentation and Registration Methods

Purpose The surgical treatment of abdominal tumors (i.e. pancreas, liver, kidney) relies on precise localization of the lesions and detailed knowledge of the patient-specific anatomy especially when a minimally invasive technique is performed. Such procedures may be improved by using a navigation system that helps the physician to identify and visualize these anatomical structures. Methods In-vitro studies were performed on custom developed phantoms that mimic both the geometric properties of the human organs, that is shape and internal structures, and the radiologic properties of the human tissue under two different imaging systems: ultrasound (US) and computed tomography (CT). Real-time US tracked, segmented and 3D reconstructed data are mapped to the pre-operative 3D model, through a fast registration procedure, assuming a locally rigid and temporarily static scenario. The segmentation process identifies the external surface of the organ and its internal structures such as duct and cysts as unstructured point clouds. By registering with the pre-operative data we have obtained from one side the deletion of the outliers in the US image and, on the other side, the update of the position, orientation and dimension of the internal structures in the CT data. Since the input data for the registration is unstructured, we have chosen a principal axes alignment algorithm based on the statistical distribution of the 3D points modeled as a mass distribution. The segmentation and registration algorithms have a straightforward interpretation, they are fast and fully automatic. Results In the validation procedure we have investigated if the accuracy and the processing time are suitable with the requirements of a navigation system for the surgical room. Ground truth was obtained using the realistic phantoms. The real-time processing of the intra-operative data was obtained by implementing the segmentation algorithm using the GPU for parallel computation. The lesions, which are the target regions in most of the clinical cases, are usually of small dimensions, compact and most of the time symmetric. Given the characteristics of these structures, the principal axes registration algorithm is not applicable, therefore the registration was used only on the external surface of the organ. We have measured the registration error both globally, considering only the external surface, and locally, on the target areas. Conclusion The use of realistic phantoms offers the possibility to identify and tune the best methods of segmentation and registration for operating room applications as well as the possibility to validate these methods. The presented framework proved to be versatile and may be applicable to a wide range of procedures. The use of registration between pre-operative and interventional data could lead to precise delineation of the tumor and therefore could increase the chances of recovery after surgery with a reduced number of side effects. The emphasis was on the real-time acquisition and processing. One limit of almost all of these procedures available in literature was the incompatibility with the real-time constraint imposed by our setup. Another limits of all the multi-modal registration solutions is the difficulty of finding a distance measure and the risk of getting stuck in a local minimum in the optimization process. The real-time solution we have chosen, based on the principal axes alignment, makes the assumption that the deformations are small because the distribution of the surface points is similar between two subsequent real-time acquisitions.