Empirical combination networks for head and neck organs at risk segmentation

To comprehensively evaluate popular medical segmentation networks on the CSTRO dataset for head and neck organs at risk (H&N OARs) segmentation, identify top performers, and integrate them into a robust hybrid network (Attention W-Net) for superior performance. U-Net, Attention U-Net, R2U-Net, UNet-plusplus, and CE-Net were selected and two novel architectures W-Net and SE-U-Net were developed. Using U-Net as the baseline, a first-stage experiment was conducted to evaluate the segmentation performance of these networks. Following initial evaluations, Attention U-Net, SE-U-Net, and W-Net achieved notably strong performance. Representative blocks were identified and extracted from these three networks to construct three hybrid architectures: Attention W-Net, SEW-Net, and Attention SEU-Net. Subsequently, a second-stage experiment was conducted to determine the optimal hybrid architecture. In the first stage, U-Net, Attention U-Net, R2U-Net, UNet-plusplus, CE-Net, W-Net and SEU-Net were tested and achieved 0.712, 0.755, 0.706, 0.710, 0.702, 0.708, 0.767, 0.749 of average dice similarity coefficient (DSC), respectively. Then the best three networks Attention U-Net, SEU-Net, W-Net were selected out. The hybrid networks Attention W-Net, Attention SEU-Net, SEW-Net were tested, and achieved 0.776, 0.768, 0.743 of average (DSC), respectively. In terms of the metric, Attention W-Net is the most effective networks for H&N OAR segmentation. The Attention W-Net and SEW-Net are the better networks which achieve better results than the popular state-of-the-arts networks for head and neck OARs segmentation.