- Abstract: Deep learning approaches such as convolutional neural networks (CNN) have achieved state-of-the-art performance in cardiac magnetic resonance (CMR) image segmentation. However, it is non-trivial to introduce shape prior knowledge to CNN-based approaches. In this paper, we employ a CNN-based method combined with image registration to develop and evaluate a shape-based bi-ventricular segmentation tool from short-axis CMR images. The method first employs a fully convolutional network (FCN) to learn the segmentation task from manually labelled ground truth CMR images generated in a low long-axis resolution. The FCN segmentation results are then used to perform a non-rigid registration using multiple high-resolution atlases, allowing the shape constraints to be inferred. This approach produces accurate, high-resolution and automatically smooth segmentation results from input images with low long-axis resolution, thus retaining clinically important global anatomical features.