AL-Net: Asymmetric Lightweight Network for Medical Image Segmentation

Medical image segmentation plays an important role in clinical applications, such as disease diagnosis and treatment planning. On the premise of ensuring segmentation accuracy, segmentation speed is also an important factor to improve diagnosis efficiency. Many medical image segmentation models based on deep learning can improve the segmentation accuracy, but ignore the model complexity and inference speed resulting in the failure of meeting the high real-time requirements of clinical applications. To address this problem, an asymmetric lightweight medical image segmentation network, namely AL-Net for short, is proposed in this paper. Firstly, AL-Net employs the pre-training RepVGG-A1 to extract rich semantic features, and reduces the channel processing to ensure the lower model complexity. Secondly, AL-Net introduces the lightweight atrous spatial pyramid pooling module as the context extractor, and combines the attention mechanism to capture the context information. Thirdly, a novel asymmetric decoder is proposed and introduced into AL-Net, which not only effectively eliminates redundant features, but also makes use of low-level features of images to improve the performance of AL-Net. Finally, the reparameterization technology is utilized in the inference stage, which effectively reduces the parameters of AL-Net and improves the inference speed of AL-Net without reducing the segmentation accuracy. The experimental results on retinal vessel, cell contour, and skin lesions segmentation datasets show that AL-Net is superior to the state-of-the-art models in terms of accuracy, parameters and inference speed.