Segmenting atrial fibrosis from late gadolinium-enhanced cardiac MRI by deep-learned features with stacked sparse auto-encoders

The late gadolinium-enhanced (LGE) MRI technique is a well-validated method for fibrosis detection in the myocardium. With this technique, the altered wash-in and wash-out contrast agent kinetics in fibrotic and healthy myocardium results in scar tissue being seen with high or enhanced signal relative to normal tissue which is ‘nulled’. Recently, great progress on LGE MRI has resulted in improved visualization of fibrosis in the left atrium (LA). This provides valuable information for treatment planning, image-based procedure guidance and clinical management in patients with atrial fibrillation (AF). Nevertheless, precise and objective atrial fibrosis segmentation (AFS) is required for accurate assessment of AF patients using LGE MRI. This is a very challenging task, not only because of the limited quality and resolution of the LGE MRI images acquired in AF but also due to the thinner wall and unpredictable morphology of the LA. Accurate and reliable segmentation of the anatomical structure of the LA myocardium is a prerequisite for accurate AFS. Most current studies rely on manual segmentation of the anatomical structures, which is very labor-intensive and subject to inter- and intra-observer variability. The subsequent AFS is normally based on unsupervised learning methods, e.g., using thresholding, histogram analysis, clustering and graph-cut based approaches, which have variable accuracy. In this study, we present a fully-automated multi-atlas propagation based whole heart segmentation method to derive the anatomical structure of the LA myocardium and pulmonary veins. This is followed by a supervised deep learning method for AFS. Twenty clinical LGE MRI scans from longstanding persistent AF patients were entered into this study retrospectively. We have demonstrated that our fully automatic method can achieve accurate and reliable AFS compared to manual delineated ground truth