Segmentation of liver from volumetric images forms the basis for surgical planning required for living donor transplantations and tumor resections surgeries. This paper introduces a novel idea of using sparse representations of liver shapes in a learned structured dictionary to produce an accurate preliminary segmentation, which is further evolved using a joint image and shape based level-set framework to obtain the final segmented volume. Structured dictionary for liver shapes can be learned from an available training dataset. The proposed approach requires only 3 orthogonal segmented masks as user-input, which is less than half the number required by current state-of-the-art interaction-based methods. The increased accuracy of the preliminary segmentation translates into faster convergence of the evolution step and highly accurate final segmentations with mean average symmetric surface distances (ASSD)  of (1.03±0.3)mm when tested on a challenging dataset containing 62 volumes. Our approach segments a volume on an average of 5 mins and, is 25% (approx.) faster than comparably performing techniques.