Automated segmentation of the left atrium from MRI cardiac images

Abnormalities of cardiac rhythm are associated with substantial morbidity and economic costs. Atrial Fibrillation is the most common cardiac electrophysiology disorder worldwide affecting around 2.3 million people just in the United States alone and is associated with an increased risk of stroke and mortality. Atrial fibrillation commonly originates in the pulmonary veins from the left atrium, a complex anatomical structure with high variation in shape and size among the population. Circumferential pulmonary veins ablation and other strategies are widely performed to treat Atrial Fibrillation. These interventional procedures can largely benefit from prior assessment of the patient LA anatomy for accurate surgical planning and intraoperative guidance. The 3D LA anatomy can be noninvasively visualized with different imaging techniques such as late gadolinium enhancement magnetic resonance imaging. In clinical practice, obtaining the complex 3D geometry of the LA is challenging even for the most experienced clinicians. Automated computer methods to segment the LA structure are highly desirable. In this work, we proposed a patch-based network methodology to segment the left atrium geometry from late gadolinium enhancement magnetic resonance images. Due to the class imbalance problem, our strategy uses what we call positive percentage parameter during training, which simply forces the network to take patches that contain at least some target segmentation. We studied the effect on the segmentation when using both region-based and contour-based loss functions with different strategies. Using a database obtained from the Department of Arrhythmias of the Hospital Clínic de Barcelona, named ClínicLA dataset, we studied the performance of our model with different hyper-parameter combinations. To compare our strategy with other approaches, we used a public LA benchmark database. We evaluated the segmentation performance in terms of Dice Similarity Coefficient and Hausdorff distance to the ground truth. Our model architecture, after the hyper-parameter fine-tuning process, enabled the segmentation of the LA in 3D with a Dice Similarity Coefficient of 88.4% in the Large ClínicLA dataset, and dice score of 91.0% in the public dataset. Both results with accurate Hausdorff distance metrics