Interactive multi-scale watershed segmentation of tumors in MR brain images

In neurosurgery, determining the position of a tumor during interventions by navigation based on preoperatively acquired (MRI) data is a common approach. In current systems that allow for navigation on preoperative data, it is assumed that no intraoperative tissue deformation occurs. However, in neurosurgical procedures, intraoperative brain deformations of 1 to 10 mm have been reported [1], making the intraoperative deformation the most important cause affecting the overall accuracy of image guided neurosurgery. Information provided by intraoperative imaging devices, which can measure the intraoperative deformation of the soft tissue, is generally less complete and less detailed. One option for solving this problem is the use of intraoperatively acquired data, e.g. ultrasound [2], to correct the preoperative images for deformations that have occurred during surgery. In order to compare tumor localization in preand intraoperative images it is necessary that the tumor is segmented in both modalities. In this paper we focus on the preoperative MR segmentation. The most common way to segment a tumor in MR images is by manual segmentation. The problem with this method is that interobserver variability can be quite large since there is no restriction in choosing the boundary of the tumor. A semi-automatic segmentation method that reduces the degrees of freedom for segmentation, while still maintaining the influence of the surgeon on the segmentation, may improve this. In this paper a semi-automatic multi-scale watershed algorithm [3,4] is evaluated for tumor segmentation in MR brain images in patients scheduled for image-guided neurosurgery, where deformation during the interventions are expected.