基于多头注意力机制的手术器械图像分割方法

ST-HRNet:Multi-head attention mechanism and high-resolution network for surgical instrument image segmentation

对手术器械的自动分割是微创手术机器人稳定运行的保障,目前的手术器械分割方法都由串联连接的,容易造成细节丢失。因此,本文提出了一种基于多头注意力机制的手术器械分割方法(ST-HRNet),采用HRNet结构构建并行子网络直接输出高分辨率特征图,防止细节丢失;还融合滑动窗口多头注意力机制来获取全局信息进一步提高分割精度。在Endovis2017手术器械数据集和私有数据集上与Unet、TransUNet、GCnet、HRnet方法进行了对比实验,实验表明ST-HRNet方法效果最佳。

The automatic segmentation of surgical instruments is the guarantee for the stable operation of minimally invasive surgical robots.The current surgical instrument segmentation methods are composed of high-to-low-resolution sub-networks connected in series,which is prone to loss of details.Therefore,this article proposes a surgical instrument segmentation method based on multi-attention mechanism and high resolution(ST-HRNet).The model uses the HRNet structure and constructs parallel sub-networks to directly output high-resolution feature maps,preventing the loss of details.In addition,the model also integrates the sliding window multi-attention mechanism to obtain global information,which further improves the accuracy of model segmentation.On the Endovis2017 surgical instrument data set and the privatedata set,the average intersection and union ratios were 95.86%and 97.17%,respectively,and the remaining indicators also exceeded the existing methods.Experiments show that STHRNet has better results than other segmentation methods.