Machine Learning Aided Analysis of Contrast-enhanced MRIs for Targeted Ablation of Atrial Fibrillation

Abstract

Atrial fibrillation (AF), characterised by rapid and irregular electrical activity in the upper chambers of the heart, is the most common sustained heart rhythm disturbance. AF is a leading cause of hospitalisation worldwide and a significant risk factor for heart failure and stroke. AF ablation procedures aim to isolate or disrupt the electrical activity sustaining AF in order to restore the normal rhythm of the heart. However, these ablation treatments remain suboptimal with limited success rates, particularly with long-term follow-up. This outcome is primarily due to a lack of basic understanding of the underlying atrial structure, particularly features of atrial geometry and fibrosis, the hallmark of structural remodelling, that directly sustain AF. Additionally, there is an absence of effective tools to create virtual patient-specific heart models and to provide accurate locations for more targeted ablation. Late gadolinium contrast-enhancing agents are used in 30% of all MRI (LGE-MRI) scans worldwide to improve the visibility of structures often associated with disease, such as fibrosis (scar tissue). Recent clinical studies using LGE-MRI suggest that atrial scar tissue is an important contributor to the structural substrate for AF, even in patients without cardiac comorbidities. The series of multi-centre, prospective, observational cohort studies from the University of Utah have demonstrated that non-invasive evaluation of left atrial (LA) fibrosis using LGE-MRI is independently associated with ablation outcomes in patients with AF. More importantly, they have developed an LGE-MRI guided patient selection approach using fibrosis extent and distribution assessment for patient stratification to either targeted AF ablation or medical management. In current medical practice, atrial segmentation from medical images for clinical diagnosis and treatment is performed by manual tracing of the atrial structures from LGE-MRIs. This process is often described as time-consuming, labour-intensive and error-prone, lacking reproducibility and scalability. However, direct automatic segmentation and analysis of the atrial structures from LGE-MRIs are challenging due to the complex atrial geometry, attenuated contrast and low spatial resolution of the in vivo LGE-MRIs, limiting the accuracy of traditional segmentation algorithms. In this thesis, we proposed and extensively tested a novel deep learning approach using a two-stage convolutional neural network architecture for direct automatic segmentation of the LA and fibrosis. First, we proposed an in-depth investigation of the conventional methods, along with state-of-the-art deep learning approaches to perform LA segmentation. This analysis extensively investigated the essential aspects of LA segmentation using deep learning, the various existing approaches and their respective shortcomings. Based on the results and insights of our investigations, we successfully developed and optimised a robust deep learning model to perform automatic LA segmentation using the largest LGE-MRIs dataset currently available (from the University of Utah). Then, we assessed the efficacy of our deep learning approach by testing it on a differential clinical dataset, translating from research development to clinical application. For this purpose, we recruited 11 patients with AF fromWaikato Hospital scheduled for AF ablation to undergo LGE-MRI. Finally, we conducted a detailed analysis of four different established methods for automatic scar tissue segmentation from LGE-MRIs, and compared the results obtained with our deep learning results when applied for LA fibrosis segmentation. This study demonstrates a promising framework to provide a 3D virtual heart including the characterisation of the key structural remodelling features implicated in AF directly from clinical LGE-MRIs. Our investigations successfully lead to the development of an efficient and robust deep learning approach for automatic LA segmentation and provide critical insights regarding the evaluation of fibrosis extent in the LA wall. This study also presents a sound basis for further investigations into new research questions regarding LA and fibrosis segmentation using a deep learning model. The conclusions from our computational framework may serve as an immediate foundation for future work on AF patient stratification, potentially patient-tailored diagnosis, ablation treatment and prognosis, to treat the most common cardiac arrhythmia seen in clinics more effectively.