A Lightweight Network with Efficient Feature Extraction for Medical Image Segmentation

Precise segmentation of medical images plays a critical role in supporting accurate clinical diagnoses. However, traditional convolutional neural networks (CNNs) often fail to capture fine edge details, leading to blurred segmentation boundaries, which makes them impractical for mobile health. To resolve these challenges, we propose a lightweight segmentation network that enhances accuracy while maintaining computational efficiency. The network introduces difference convolution which integrates prior knowledge into convolutional layers to enhance representation and generalization, enabling efficient feature extraction. We employ Spatial and Channel Synergistic Attention to enhance the integration of spatial details and semantic representations across the encoder–decoder pathway. A six-stage encoder-decoder structure with deep supervision further enhances edge representation and ensures global semantic consistency. With a parameter size of only 17KB, the model is 486x smaller and 574x more computationally efficient than the traditional U-Net. Extensive evaluations show that our model surpasses existing approaches in accuracy and efficiency, making it a viable solution for real-world medical image segmentation applications.