M. Linder. Image Classification Approaches for Discrimination of Different Atrial Flutter Mechanisms. Institute of Biomedical Engineering (IBT), Karlsruhe Institute of Technology (KIT). Masterarbeit. 2020
Atrial flutter (AFL) is a common heart rhythm disorder, which is characterised by regularly propagating electrical signals and self-sustained electrophysiological mechanisms. Since AFl mechanisms are usually discriminated from invasive intra-cardiac signals, enabling a non-invasive categorisation of driving mechanisms prior to the invasive procedure, would greatly benefit the ablation strategy. In the present work, various image and video classifi- cation approaches are implemented and evaluated, in order to discriminate distinguishable electrophysiological mechanisms sustaining AFl. Therefore, the cardiac excitation of 20 different AFl scenarios, was simulated on eight volumetric bi-atrial anatomies. Solving the forward problem of electrocardiography, the body surface potential maps (BSPMs) were calculated and projected onto eight triangulated torso meshes. Given by the implementation of 124 atrial rotations within the different torso models, the provided database comprising 153,760 simulations, was completed. Finally, with further processing the generated 12-lead electrocardiograms (ECGs) and spatially down- sampling the obtained BSPMs, a consistent ground truth for AFl perpetuation mechanisms, was established. In the first part of this work, the resulting recurrence plots (RPs) and distance plots (DPs) were discriminated via the atlas and K-nearest-neighbour (KNN) classification approach as well as referring to two residual convolutional neural networks (CNNs). Moreover, for each approach the ascription into 20 AFl mechanisms and five macro groups was performed independently. Hit rates of 21% considering the assignment into 20 AFl scenarios and 40% corresponding to the macro groups, were achieved. Focused on the classification of the BSPM video sequences, the second part of this work contained the assignments based on a three-dimensional residual network (ResNet). This resulted in average accuracy values of 57%, in cases of 20 AFl mechanisms and 67% considering the simplified macro groups ascription task. Thus, the BSPM-based categorisation has been shown an effective method to discriminate distinguishable AFl electrophysiological mechanisms in a non-invasive in silico study. It is outlined that this could help to delineate the ablation strategy, reduce resources to conduct invasive cardiac mapping and time.