Registration of Real-Time and Prior Images for MRI-Guided Cardiac Interventions.

摘要

Cardiac magnetic resonance imaging (MRI) is a promising modality for guiding interventions within the heart. Compared to the gold standard of x-ray fluoroscopy, MRI provides superior soft tissue contrast for visualizing the cardiac anatomy without the use of ionizing radiation. One particular application of interest is the use of MRI to guide an interventional catheter into the left ventricle (LV), and applying radio-frequency (RF) energy to eliminate the anatomical substrates responsible for ventricular tachycardia. However, interventional guidance with MRI is not without its own unique set of challenges. Different MRI acquisition schemes have tradeoffs in terms of image quality and acquisition time. High quality 3D roadmap volumes can be acquired prior to the procedure, but accurate guidance with this dataset can be limited during the procedure due to dynamic motions of the heart. Alternatively, fast 2D real-time acquisition schemes can be used to rapidly acquire images that reflect the dynamic motions of the heart at the expense of imaging quality and spatial coverage. Unfortunately, neither approach is ideal for guiding an interventional catheter to an appropriate target location.;Therefore, this thesis presents and tests novel ideas for combining the advantages of motion characterization from the real-time images and accurate visualization of anatomical structures from the pre-operative images. To this end, an image-based registration algorithm for aligning the prior roadmap image with the real-time images was developed first. This proposed algorithm synchronized the cardiac phase from both datasets, and corrected for respiratory motion induced misalignment between the two datasets. Promising results were obtained in a volunteer cohort, but the limitation of the algorithm was its extensive computational complexity, which made it infeasible for real-time correction during the procedure. To solve this problem, a motion modeling approach was designed to correct for motion-induced misalignment between the prior and real-time images. In this approach, the initial derivation of the motion model would still be time consuming, but the application of the motion model based correction would be very fast. In a pre-clinical RF ablation study, it was successfully demonstrated that the motion modeling approach can be used to improve the accuracy of ablation targeting. However, one limitation of the motion model was that it is derived based on an initial set of calibration images, which may not accurately reflect the motions of the heart over an extended period of time. Therefore, an improvement to the motion modeling approach was pro- posed via implementation of a graphic processing unit (GPU) accelerated algorithm to dynamically update the motion model in real-time. The dynamic motion model could be continuously updated during the entire procedure, to account for respiratory drift and changes in breathing pattern. The proposed technique was applied to a volunteer cohort, which showed that the GPU accelerated dynamic motion model provided more accurate motion correction over time compared to previous approaches.;The methods presented in this thesis are intended to improve the accuracy of MRI- guided cardiac interventional procedures. In the final GPU-based motion correction method, it was demonstrated that the desired registration accuracy (i.e., < 5 mm) could be achieved, and the adaptive motion model could be updated in real-time (i.e., < 1 s). Further studies are proposed to apply the described motion correction tools to prospectively guide an RF ablation study in preclinical subjects.