With the progress of medical imaging technology, accurate segmentation of cardiac images plays a vital role in diagnosing and treating related diseases. However, the presence of boundary ambiguity and intensity inhomogeneity in cardiac MRI images limits the accuracy of segmentation results. To address this issue, we introduce a self-supervised Few-shot Semantic Segmentation (FSS) framework for cardiac image segmentation. This framework incorporates masked image modeling and class-level prototype pooling module to enhance segmentation accuracy. Our model achieves the Dice score of 73.9% on ACDC dataset for 1-shot setting, surpassing existing techniques by 6.1%.