Markov chain modeling of ECG-gated live left atrial fluoroscopy variability to establish a well-defined basis for rigid registration to a three-dimensional CT image
Abstract
Real-time 2-dimensional X-ray to acquired 3-dimensional computed tomography (CT) image registration is currently of interest for improving visualization in the fluoroscopy guided catheter ablation treatment of atrial fibrillation. An important feature of this registration is 3d pose estimation of the left atrium from the 2d fluoroscopy image prior to registration. Computational complexity constraints on this real-time application limit registration to rigid methods. Aimed at satisfying the rigidity assumption, ECG gating is employed to acquire images at fixed phases of the cardiac cycle to circumvent the otherwise continuous elastic deformations of the cardiac chamber. Observations of the ECG gated fluoroscopy sequences, however, yield dynamic variability in the location of registration landmark features across the image sequences. There is currently no protocol for identifying which gated fluoroscopy frame to use in the rigid registration to the CT image. As such, the registration process is not sufficiently well-defined to address the issue of 3d pose estimation. A standard protocol for establishing a well-defined registration representative from the fluoroscopy images is therefore desired. This thesis presents a novel Markov chain method as such a protocol. In this method, patient specific Markov chains are identified from patient data. Empirical transition matrices and the associated unique limit distributions are defined and used to identify a set of registration points. The method was tested on sequences of patient ECG gated left atrial fluoroscopes. A notion of optimality in a set of representative registration points is defined and optimality measures designed to quantify these components were computed and compared to points identified by two control methods. The results indicated that the MC identified representative points converged rapidly to a stable set once a threshold level of input sequence length was reached. Comparison with the control methods indicated that the MC method was an improvement in each of the optimality measures over the existing random approach. Additionally, the MC method showed optimal stability over the other methods with respect to longer data sequences. This has positive implications for the ablation procedure that follows registration. The well-defined registration representatives form a rigid basis for addressing the challenge of 3d pose estimation from the fluoroscopy images. This is discussed in the context of ongoing and future work.
This paper has been withdrawn.