Evaluation with an Independent Dataset of a Deep Learning-based Left Atrium Segmentation Method

Abstract

Atrial fibrillation (AF) is the most prevalent type of arrhythmia nowadays. Even though it is associated with significant morbidity and mortality, there is still a substantial lack of basic understanding of the left atrium (LA) and pulmonary veins (PVs) anatomical structure that curbs the performance of current clinical treatments for the disease. Thus, segmentation and 3D reconstruction of the LA and PVs are of crucial importance for the diagnosis and treatment of AF. In this context, cardiac 3D Late Gadolinium Magnetic Resonance Imaging (LGE-MRI) appear as a very good tool for cardiac tissue characterization and myocardial fibrosis detection. In fact, these images have been proofed as reliable predictors of catheter ablation success, which is often the chosen treatment for AF patients. Several manual and semi-automatic segmentation tools from LGE-MRI scans are currently in use, but these are very time-consuming and highly prone to errors, hence the need for an automatic segmentation approach. With the rise of deep learning and convolutional neural networks, a number of automatic schemes are being developed. In this project, we evaluate a model that has been developed at the Hospital Clínic de Barcelona for obtaining an automatic segmentation of the LA using a deep learning architecture. Concretely, we tested this model with an independent set of images from another MRI vendor, and we obtained a set of quantitative and qualitative measures to validate the results. For the pursuit of our aims, this work begins with the state-of-the-art for LA segmentation of LGEMRI scans and with a market analysis of the field. We then present our proposed solution together with the obtained results and the corresponding conclusions.