Machine Learning- and Model-based Image Fusion for cardiac interventions

Student thesis: Doctoral ThesisDoctor of Philosophy


In cardiac interventions, such as cardiac resynchronization therapy (CRT), fluoroscopy guidance can be enhanced through overlaying information extracted from preoperative magnetic resonance imaging (MRI) data. Registering multimodal image data, such as 3D/2D cine MRI to X-ray, however, remains a major research challenge. Due to fundamental differences in the image acquisition physics of MRI and X-ray fluoroscopy, no similar intensities or shared features are available. Due to the lack of shared features between the modalities, the use of classical intensity- or feature-based registration approaches is not feasible. This thesis proposes two main approaches to address this problem: 1) adjacent anatomical model-based registration and 2) imitation learning-based model-to-image registration. The adjacent anatomy-based approach relies on extracted models of the left ventricle (LV) from MRI and a reconstructed point cloud of the coronary veins from two interventional, contrasted X-rays (venograms). The method exploits the anatomical adjacency of the LV and the coronary veins through a globally optimal point cloud registration: globally optimal iterative closest point (GO-ICP). The approach has demonstrated high robustness and accuracy on phantom and clinical patient data. However, the approach is greatly dependent on the quality of acquired venograms. To create a more generic approach, an imitation learning-based method is proposed that is able to register the LV model to a single, non-contrast-enhanced X-ray acquisition. An artificial neural network (ANN) predicts transformations of the 3D LV model iteratively, relative to the 2D X-ray image to register to. The training of the ANN is performed entirely on digitally reconstructed radiographs (DRRs), artificial X-ray data generated from computed tomography (CT) volumes. The approach provides high accuracy on DRRs and high robustness on clinical X-ray data, however, results on the target domain are highly training parameter-dependent. The influence of parameters, i.e., training data order and network weight initialization, is investigated extensively. It is shown that domain randomization, i.e., applying unrealistic perturbations to the synthetic training data, can result in substantially more consistent robustness on the target domain. The imitation learning-based registration approach was integrated into a clinical prototype for the interventional guidance of CRT procedures. After initial, successfully guided cases, the approach is to be extensively validated in a multicenter clinical trial.
Date of Award1 Jul 2019
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorKawal Rhode (Supervisor)

Cite this