Cardiac magnetic resonance imaging has been proven to be a great aid tool in clinical diagnosis. Computational models arising from these images have been developed for many years by engineers, radiologists and clinicians. A first task in this process is to segment the different regions of the heart, where machine learning and, more recently, deep learning, have shown good performance. My project aims to improve the current network performance when segmenting the left-ventricular, myocardial and right-ventricular regions through (1) data augmentation, (2) data-set combination and (3) loss-function optimization, with a limited amount of computational resources. Results show improvements for all three methodologies. In addition, investing computational resources on muscular regions provides better performance in cavity regions.