Registration Uncertainty in Image Guided Therapy

The main goal of a neurosurgical tumor resection is to maximize the removal of tumor tissue while conserving healthy and important functional brain tissue. Functional areas are commonly located in the pre-operative image space and a non-rigid registration used to map this information into the intra-operative space to provide the surgeon with accurate intra-operative information regarding the boundaries between functional and tumor tissue. Admittedly, neurosurgery is a relatively low-volume procedure. His torically, however, the feasibility and significance of novel image guided techniques are initially proven in neurosurgical applications because of the simplifying constraints the rigid skull provides. One high-volume application where, once developed fur ther, this framework could have significant impact, is in radiation oncology treatment. We propose to develop a probabilistic non-rigid image registration framework that will reduce the risks involved in image guided therapy in general and neurosurgical t umor resections in particular. We formulate the registration problem in a Bayesian framework and include a non-linear bio-mechanical deformation model which accommodates resection. The full posterior distribution on deformations will be characterized by w ay of MCMC sampling techniques. MCMC methods are in general slow. We will use large scale computational clusters together with parallel Markov chains to speed up computations. From the posterior distribution we can find the most likely deformation as well as the uncertainty of the estimated deformation parameters. Providing the surgeon with registration uncertainty estimates will increase the surgeon's confidence level and his ability to make better intra-operative decisions. We will integrate the develop ed methods in an open-source free software package called 3D Slicer which we will use to evaluate the impact of registration uncertainty on neurosurgical decision making.