Atrial fibrillation (AF) is the most common arrhythmia and is associated with substantial morbidity and mortality. However, the current treatment of AF is suboptimal, partially due to our lack of basic understanding of the underlying atrial substrate which sustains AF directly in human atria and need of quantitative tools to investigate the optimal ablation strategy in clinical/experimental settings. Accurate analysis of 3D atrial anatomy and its underlying structure from medical images, particularly late gadolinium enhancement magnetic resonance imaging (LGE-MRI), provides an effective patient-specific approach for clinical diagnosis and targeted treatment for patients with AF. In this talk, I will report our recent research development based on late gadolinium enhancement MRI from patients with AF, ranging from robust automatic 3D atrial segmentation using a novel machine learning approach to critical insights obtained using the structure-and-function detailed computer models.