Lightweight Liver Image Segmentation Algorithm Based on MSG-Unet

In this paper, we propose a lightweight U-Net-based liver image segmentation model, MSG-UNet, which aims to improve the segmentation accuracy while reducing the computational complexity to meet the application requirements in the resource-constrained environments. The model adopted MobileNetV4 to replace the encoder of U-Net, and utilized its techniques such as depth-separable convolution to significantly reduce the computation while maintaining the feature extraction capability. GhostConv is introduced in the decoder to replace the ordinary convolution, which generates redundant feature maps through simple linear transformations, significantly reducing the computational complexity and the number of parameters. In addition, the SE attention mechanism is embedded in the jump connection to optimize the feature information transfer, highlight the liver-related features, suppress the useless features, and improve a segmentation effect. The model is experimented on LiTS17 dataset, and the DSC and IoU coefficients of the model are 91.79 % and 86.77 %, respectively, and the computational and parametric quantities are 5.85 GFLOPs and 4.66 M, respectively.Compared with the traditional U-Net and other lightweight models, MSG-UNet significantly reduces the number of model parametric quantities and computational quantities, while maintaining high segmentation accuracy.