基于注意力机制的U-Net网络模型分割X线图像椎弓根影

U-NET NETWORK MODEL BASED ON ATTENTION MECHANISM FOR SEGMENTING PEDICLES OF VERTEBRAL ARCH IN RADIOGRAPHS

脊椎X线图像中椎弓根影面积小且对比度不高,现有的X线图像椎弓根影的分割方法存在精度较低且效率不高的问题,对此提出一种端到端的基于注意力机制的U-Net神经网络的脊椎X线图像椎弓根影的分割方法。在原始U-Net网络中嵌入通道注意力模块和注意力门模块,既提高了网络对目标区域提取的准确性又解决了模型跳跃连接的冗余问题,从而实现高精度全自动的X线图像椎弓根影分割。实验结果表明,采用改进后的U-Net网络模型分割X线图像椎弓根影平均准确率为99.13%,平均Dice系数为89.01%,平均精确度为88.75%,平均召回率为89.80%,平均Hausdorff距离为3.8316像素,优于现有的自动化分割方法。

In vertebral radiographs,the pedicle of vertebral arch area is small and the contrast is not high.The existing methods of segmenting pedicles of vertebral arch radiographs have problems of low accuracy and inefficiency.To this end,an end-to-end attention-based U-Net neural network-based segmentation method for pedicles of vertebral arch radiographs is proposed.The Channel Attention module and Attention Gates module were embedded in the original U-Net network to improve the accuracy of the original network's extraction of the target area while solving the redundancy problem of model jump connections,so as to achieve a high-precision,automatic vertebral radiographs pedicles segmentation.Experimental results show that the improved U-Net network model is better than the existing automated segmentation methods,with an average accuracy of 99.13%,an average Dice coefficient of 89.01%,an average precision of 88.75%,an average recall rate of 89.80%and an average Hausdorff distance of 3.8316 pixels.