Modeling Prior Knowledge for Image Registration in Liver Surgery

The careful consideration of blood vessels and the complete removal of the tumor are essential in oncological liver surgery to preserve healthy liver tissue and to minimize the probability for recurrent tumors. The enormous improvements in medical imaging over the last 20 years enable an accurate computer assisted 3D planning of the surgical intervention. The accurate transfer of the preoperative plan to the patient on the operating table is not trivial as the liver deforms due to intraoperative bedding and mobilization of the organ. Intraoperative 3D ultrasound is a possibility to capture the current shape and position of the liver during a surgical intervention. In the 3D ultrasound volume a navigation system shows the accurate position of the surgical instrument and its spatial relation to the vessels and the tumor. The key problem for the transfer of the surgical plan is the compensation of the deformations between preoperative images resp. planning models and the intraoperative ultrasound data. Such problems have not yet been solved satisfactory. The image processing technique to compensate this is called nonrigid registration. Non-rigid registration is also needed for the postoperative control based on a comparison between pre- and postoperative images. The principle difficulty of non-rigid registration is the vast number of theoretically possible non-rigid transformations, of which only a small subset compensates the present anatomical deformations. The fundamental hypothesis, pursued by this thesis, is that the incorporation of a priori knowledge about the image contents or about application-specific transformation properties significantly reduces the number of admissible transformations. We develop a new distance measure which considers the tube-like shapes of vessels by specific local filters, which give high responses, if the preoperative vessel models fit the appearance of vessels at the same position in the intraoperative image. A priori knowledge about anatomical corresponding landmarks is a direct restriction on the transformation. An important property, which sets our method apart from previous work, is that anisotropic tolerances to compensate landmark localization uncertainties are consequently integrated into pure landmark schemes as well as into schemes combining intensity and landmark information. The developed registration methods are validated on clinical image data by a new reference standard.