CASNet is a new computer method for automatic segmentation of cardiac structures in magnetic resonance imaging (MRI). The method is based on the existing U-Net architecture and includes three main improvements: Multi-Scale Context Block, which processes cardiac structures in different sizes, Cross-Attentive Skip Connections, allowing selective selection of important features in image processing, and curvature-focused loss, which improves smoothness and anatomical accuracy of segmentation boundaries. Accurate segmentation of cardiac structures is essential for reliable diagnosis and analysis of cardiovascular diseases. Conventional neural networks struggle to maintain proper semantic consistency and geometric smoothness, especially in slices with high anatomical variability. Tests on the ACDC dataset showed that CASNet outperforms basic U-Net models and recent attention-based architectures in the region overlap and boundary accuracy metrics. The proposed approach provides a robust solution for high-precision cardiac MRI segmentation with potential for future clinical applications in AI-assisted cardiac analysis.