Late Gadolinium Enhancement Magnetic Resonance Imaging (LGE MRI) emerged as a routine scan for patients with atrial fibrillation (AF). However, due to the low image quality automating the quantification and analysis of the atrial scars is challenging. In this study, we pro-posed a fully automated method based on the graph-cuts framework, where the potential of the graph is learned on a surface mesh of the left atrium (LA) using an equidistant projection and a Deep Neural Network (DNN). For validation, we employed 100 datasets with manual delineation. The results showed that the performance of the proposed method improved and converged with respect to the increased size of training patches, which provide important features of the structural and texture information learned by the DNN. The segmentation could be further improved when the contribution from the t-link and n-link is balanced, thanks to inter-relationship learned by the DNN for the graph-cuts algorithm. Compared with the published methods which mostly acquired manual delineation of the LA or LA wall, our method is fully automatic and demonstrated evidently better results with statistical significance. Finally, the accuracy of quantifying the scars assessed by the Dice score was 0.570. The results are promising and the method can be useful in diagnosis and prognosis of AF.