LAFU-Net: Lightweight Left Atrium Segmentation Network Based on U-Shaped Network

Left atrial segmentation is critical for the diagnosis and treatment of atrial fibrillation through image analysis. Traditional left atrial segmentation models are often computationally complex and resource-intensive, making them unsuitable for resource-constrained clinical environments. Therefore, developing lightweight medical image segmentation models has become essential. Lightweight models enhance efficiency and real-time performance by reducing parameters and computational load, making them more suitable for smart healthcare applications on embedded devices and mobile platforms. While U-Net and its derivatives have been widely applied in medical image segmentation tasks, they struggle to meet the demands of mobile healthcare due to their large number of parameters, computational complexity, and slow inference speeds. To address these challenges, this study proposes a lightweight left atrial segmentation network, LAFU-Net, which aims to strike an optimal balance between model size and segmentation accuracy. First, LAFU-Net introduces the Mixed Dilated Convolution Unit (MDCU), which modulates the dilation rate of the convolutional kernel. This mechanism expands the receptive field without increasing the number of parameters or computational cost, thus balancing model complexity reduction with improved computational efficiency. Second, considering the limitations of existing attention modules like the Convolutional Block Attention Module (CBAM) in resource-constrained environments, we propose an enhanced version-Enhanced Convolutional Block Attention Module (ECBAM). ECBAM adaptively adjusts the model's attention to key features by incorporating spatial positional encoding and memory storage, thereby enhancing the representation of essential features while suppressing irrelevant information. Finally, the entire LAFU-Net architecture is designed with a lightweight structure to minimize redundancy and improve computational efficiency. Experimental results demonstrate that, compared to existing methods, LAFU-Net achieves superior segmentation performance while maintaining a significantly lower number of parameters, making it well-suited for application in resource-constrained environments.