Neurovascular pathologies are often treated with the help of imaging to guide catheters inside arteries. However, positioning a micro-catheter into the aortic arch and threading it through blood vessels for embolization, mechanical thrombectomy or stenting is a challenging task. Indeed, adverse aortic arch anatomies are frequently encountered, especially when the aortic arch is dilated, or the supra-aortic branches are elongated and tortuous. In this article, we propose a pipeline using convolutional neural networks for the segmentation of the aortic arch from magnetic resonance images for further anatomy classification purpose. This pipeline is composed of two successive modules, dedicated to the localization and the accurate segmentation of the aortic arch and the origin of supra-aortic branches, respectively. These segmentations are then used to generate 3D models from which the anatomy and the type of the aortic arches can be characterized. A quantitative evaluation of this approach, carried out on various U-Net architectures and different optimizers, leads to satisfactory segmentation results, then allowing a reliable characterization.