Fast 3D Model Update for the Compensation of Soft Tissue Deformation in Image Guided Liver Surgery

Today, the use of computer assistance to guide surgeon's gestures becomes increasingly common for interventions on rigid organs such as the hips and vertebrae. Guidance information is provided through the representation of the position of surgical instruments on the patient's image data, acquired pre-operatively from MRI or CT scanners, or on ultrasound images acquired intra-operatively. Furthermore, raw image data are often fused with virtual three-dimensional reconstructions (3D models) of anatomical structures, which have been created during planning, to improve the surgeon's orientation and ease localization. However, at present, these techniques are applicable only in cases where the organ can be mobilized and its shape does not vary with time. In the case of liver resection, the medical context of this thesis, it is crucial to accurately bound the area occupied by a tumor mass, in order to completely remove it and minimize the risk of metastases proliferation. Moreover, it is important to perform the resection along specific planes, to avoid damaging the remaining vessels and allow the liver to regenerate and continuously ensure its vital functions. Therefore, using navigation systems to relate data of the pre-operative planning, such as resection planes, with intra-operative ultrasound images, reveals its usefulness. Unfortunately, with soft tissues, the fusion of pre- and intra-operative data remains highly problematic. In fact, the dynamic registration of data sets is much more difficult to perform, due to the large deformations that may affect the organs, such as those of the abdominal cavity or pelvis. The objective of this thesis is to demonstrate that mass-spring based deformable models, those commonly used in surgical simulators, can be used to update, during surgery, the shape and position of the 3D model of liver vessels, which were reconstructed before the operation. The process must be fast in order to satisfy surgery time constraints. To this end, we have developed a new type of non-rigid registration based on such deformable models. The network of masses and springs is located on the 3D model centerline and deforms under the influence of external constraints. These constraints are based on the distance between that deformable skeleton and a set of points which represent the intra-operative position of the vessels. The point position is set from the automatic segmentation of the veins in ultrasound images. An experimental setup based on a porcine liver under isolated perfusion was created to evaluate the efficiency of the registration and segmentation methods in non-simulated conditions. The results show that by using intra-operative measurements with reasonable noise level, it is possible to rapidly update the 3D preoperative models, providing powerful guidance information to surgeons. In addition, measurements made on human patients allowed us to analyze in details the source of errors and suggest possible improvements of the method. Finally, we introduce a probabilistic approach, based on particle filters, to determine experimentally the spring stiffness parameters of our deformable models. The results obtained by simulation demonstrate the potential of this approach in navigation systems for surgery, combining force measurements and intra-operative imaging.