Accurate delineation of cardiac structures from echocardiographic (echo) images is crucial for quantifying clin- ical indices such as ejection fraction and chamber volumes. However, echo images are challenging due to speckle noise, shadowing, low contrast boundaries, and inter-patient variability. This paper presents a lightweight yet effective segmentation framework that integrates a compact 3-layer Residual Network (ResNet-3) encoder with a spatial channel attention mechanism to enhance boundary localization and suppress artifacts. The model is designed to achieve strong performance with reduced parameters, making it suitable for real-time clinical workflows. Experiments on a cardiac echo segmentation dataset demonstrate improved Dice and IoU over U-Net and attention free residual baselines, with notable gains on difficult frames exhibiting weak endocardial borders. The proposed approach provides a practical balance between accuracy and efficiency for robust cardiac structure segmentation.

Index Terms: Echocardiography, cardiac segmentation, ResNet, attention mechanism, deep learning, U-Net, medical imaging