Date of Award
Doctor of Philosophy (PhD)
Mathematics, Statistics and Computer Science
An Electrophysiology study is conducted to diagnose and treat heart rhythm disorders, such as arrhythmias (abnormal heartbeat) like atrial fibrillation. A catheter is inserted into the chamber of interest to acquire 3D location and electrical information to create an electroanatomical map. This dissertation explores the design of a mapping system based on interpretable generative neural networks for generating patient specific cardiac models. Chapter 1 provides an introduction to electroanatomical mapping, the need for interpretability in neural networks and other relevant topics that are discussed in detail in the chapters that follow. Neural networks are often very large models with millions of parameters and can be difficult to train or draw inferences on compute constrained devices. Chapter 2 explores a formal principled Bayesian technique to eliminate parameters (connections) in a neural network. Prior information about the “weights” of the connections is quantified in the form of prior distributions, combined with data likelihoods, to yield a formal posterior distribution for the parameters of the model. From this posterior distribution, formal hypothesis tests are performed to eliminate connections. This makes the neural network smaller, simpler, and more explainable. Chapters 3 and 4 explores how approximate Bayesian inference can be utilized to model structures in volumetric data (CT/MRI). We explore how a full Bayesian approach can quantify uncertainty and help improve interpretability in results generated by neural network models. In Chapter 5 we build an electroanatomical mapping system based on the frameworks developed in the previous chapters that is capable of generating patient specific cardiac chamber models that are interpretable and useful for navigation in Electrophysiology studies.
Available for download on Thursday, August 01, 2024