Mesh-derived image partition for 3D-2D registration in image-guided interventions

Image-guided interventions conducted under a 2D modality benefit from the overlay of relevant 3D information from the preoperative stage. The enabling technology for this overlay is 3D-2D registration: the process of finding the spatial pose of a 3D preoperative image in relation to 2D intraoperative images. The successful integration of a registration solution to a surgery has the potential for significant positive impact in terms of likelihood of treatment success and intervention duration. However, many surgeries are routinely done without the assistance of registration because no current solution is practical in their clinical context. In order to remedy these issues, we focus on producing practical, targeted registration solutions to assist image-guided interventions.;The first surgery we address is catheter ablation for atrial fibrillation (CA for AF), an electrophysiology procedure to treat heart arrhythmia conducted under X-ray fluoroscopy. In this surgery, a 3D image, either magnetic resonance (MR) or computed tomography (CT), is taken preoperatively to define the anatomy of the left atrium (LA) and pulmonary veins (PV)s. A mesh, segmented from the 3D image, is subsequently used to help positioning the ablation catheter via its overlay on the intraoperative fluoroscopic images. Current clinical registration solutions for CA for AF are slow and often require extensive manual manipulations such as the identification of fiducial points, which is problematic when intraoperative updates of the 3D image's pose are required because of patient movement. The automatic solutions are currently not precise enough to be used clinically. Also, the solutions which do not involve electroanatomic mapping are not suitable for MR/fluoroscopy registration. This is problematic since we target CA for AF interventions where the 3D modality is MR and electroanatomic mapping is not used.;There are two principal challenges to overcome in order to provide a clinically useful registration algorithm. First, solving the notoriously hard MR to X-ray fluoroscopy registration problem which is further complicated in cases of CA for AF because of the partial match between modalities at the level of the PVs. Second, solving the registration quickly enough to allow for intraoperative updates required due to the patient's movement. We introduce a new registration methodology based on mesh-derived image partition (MDIP) which uses projections of a mesh segmented from the 3D image in order to infer a segmentation of the 2D X-ray fluoroscopy images. This is orders of magnitude faster than producing volumetric projections and, since the mesh can be segmented from either MR or CT, the same procedure is valid for both modalities. The fitness of the registration is evaluated by custom-built similarity measures that compare the statistical properties of the segmented zones and incorporates mask-depth information to account for the partial match at the level of the PVs.;We validate the MDIP algorithm on 7 cases of patients undergoing CA for AF surgery. Four MDIP-based similarity measures are introduced; each one is validated on 1400 biplane registrations. The precision, range, speed and robustness of the solution is assessed by calculating the distribution of projection distance error in function of the correctness of the initial pose for all 5600 biplane registrations. The precision is also evaluated visually by overlaying the ground-truths with results from the registration algorithm. To give a fair appraisal of the expected behavior, the examples are taken from cases exemplifying the average error measured as well as one standard deviation above and under.;The registration algorithm is also applied to cases of sclerotherapy for venous malformation (SfVM) in order to assess its portability to other type of surgeries. SfVM are especially challenging because the malformation can be present on any body part, which offers little predictability on the properties of the medical images from one patient to another. Our dataset is sampled from monoplane surgeries and did not come with metadata allowing a geometrical calibration of the system. We demonstrate that MDIP-based registration is applicable to cases of monoplane SfVM, but that modifications are required in order to account for the wide variety of body parts where VMs are common. The protocol developed for CA for AF surgeries can be used for embolizations or when the interior/exterior border of the organ is prominent, but gradient information has to be taken into account by the similarity measures in order to properly register cases where bones are predominant.