Medical image segmentation plays a vital role in intelligent healthcare, with the U-Net architecture widely recognized for its strong performance. However, many U-Net variants remain impractical for mobile and real-time applications due to their large parameter counts and high computational demands, and they often underperform in multiscale feature representation and generalization. To overcome these limitations, we introduce EL-UNet, a lightweight model that preserves U-Net’s architectural strengths while improving both efficiency and accuracy. EL-UNet incorporates an Attentive Multi-Scale Block (AMSB), which unites multiscale convolutions with channel and spatial attention, to better capture critical features of small structures and complex backgrounds. A novel gated skip connection further mitigates semantic discrepancies between encoder and decoder features. By adopting depthwise separable convolutions and a streamlined channel design, EL-UNet substantially reduces the parameter count and computational overhead. Experiments on multiple public datasets demonstrate that EL-UNet outperforms recent U-Net variants in precision, mIoU, and Dice metrics, achieving an optimal balance between resource efficiency and segmentation performance.