In this work, we implement a fully convolutional segmenter featuring both a learned group structure and a regularized weight-pruner to reduce the high computational cost in volumetric image segmentation. We validated our framework on the ACDC dataset featuring one healthy and four pathology groups imaged throughout the cardiac cycle. Our technique achieved Dice scores of 96.8% (LV blood-pool), 93.3% (RV blood-pool) and 90.0% (LV Myocardium) with five-fold cross-validation and yielded similar clinical parameters as those estimated from the ground truth segmentation data. Based on these results, this technique has the potential to become an efficient and competitive cardiac image segmentation tool that may be used for cardiac computer-aided diagnosis, planning, and guidance applications.