Accurate sperm segmentation is essential for automated semen analysis and male fertility assessment, yet existing methods struggle to balance segmentation accuracy with computational efficiency. We propose Dynamic U-Net (Dy-UNet), a novel lightweight deep learning architecture that achieves state-of-the-art performance while maintaining exceptional efficiency. The architecture integrates four key innovations: a cross-resolution dual-branch network that preserves high-resolution spatial details, a dynamic snake convolution that adapts to curved sperm morphology, a grouped multi-axis Hadamard product attention that efficiently captures long-range dependencies, and a fast cross-resolution aggregation that enables multi-scale feature fusion. Comprehensive evaluation on the SegSperm dataset demonstrates that Dy-UNet substantially outperforms existing methods. Dy-UNet achieves a state-of-the-art mean Intersection over Union of 0.44260 and Dice similarity coefficient of 0.61362, which surpasses the baseline U-Net by 25.5% and 17.7%, respectively. Remarkably, Dy-UNet achieves this superior performance with only 184 K parameters, especially suitable for deploying mobile platforms and embedded systems in clinical settings. The lightweight architecture enables accurate, objective, and accessible automated sperm morphology assessment with significant implications for clinical fertility diagnosis and point-of-care testing.