A lightweight memory-efficient 3D U-Net with attention and uncertainty modeling for coronary artery segmentation

Deep learning (DL) methods have been widely applied to 3D coronary artery segmentation. However, due to the large data volume of 3D CT images, efficient GPU memory utilization has become a major bottleneck for the practical deployment of current DL segmentation models. In addition, effectively capturing tubular structure information while suppressing background noise in feature maps remains a significant challenge. To address these issues, we propose a Memory-Efficient U-Net (ME-UNet)—a lightweight U-Net variant that removes residual connections in each block and introduces two key components: the Uncertainty Skip Connection (USC) and Dimensional Fusion Attention (DFA). Experimental results show that ME-UNet not only reduces GPU memory consumption but also achieves state-of-the-art segmentation performance. The source code is available at https://github.com/hd1437/ME-UNet . • We propose ME-UNet, a lightweight, memory-efficient 3D CCTA vessel segmentation model. • Dimensional Fusion Attention for efficient global–local context modeling is introduced. • Uncertainty Skip Connections suppress noise and enhance robustness. • On ImageCAS, ME-UNet surpasses UNet++ in AUC (+0.011), DSC (+0.007), IoU (+0.013). • GPU memory is reduced to ∼ 4.7 GB vs 18.4 GB (UNet++), with favorable runtime characteristics.