基于改进U-net++的超声图像结直肠子宫内膜异位区域检测

Improved U-net++ Network for detection of intestinal endometriosis based on ultrasound image

探究基于改进U-net++网络以及增加多通道特征融合的方法,实现准确高效的超声图像结直肠子宫内膜异位区域自动检测。所提神经网络在U-net++为原型的分割网络上进行改进,采用端到端的结构,输入超声图像和其边缘提取图像,输出结直肠子宫内膜异位区域检测结果。实验数据来自深圳市人民医院的166例结直肠子宫内膜异位患者的超声内镜图像,随机选择133例作为训练样本,33例测试样本。在网络训练过程中,采用十折交叉验证法进行验证。结果说明:在33例测试集样本上,方法最终的平均检出率、精确率、召回率分别为90.9%、72.4%、89.8%。改进神经网络以及多通道特征融合输入的方式可自动检测结直肠子宫内膜异位区域,且检测鲁棒性及精度较高,可作为参考辅助医生进行临床决策和干预。

Explore the method of improving the U-net++ network and adding multi-channel feature fusion based on ultrasound images to achieve accurate and efficient automatic detection of intestinal endometriosis. The proposed neural network is improved on the segmentation network based on U-net++, adopting an end-to-end structure, inputting ultrasound image and its edge extraction image, and outputting the detection result of intestinal endometriosis.The experimental data were obtained from the endoscopic endoscopy images of 166 patients with intestinal endometriosis in Shenzhen People’s Hospital. 133 cases were randomly selected as training samples and 33 test samples.In the network training process, the ten-fold cross-validation method is used for verification. The results show that On the 33 test set samples, the final average detection rate, precision rate, and recall rate of this method were 90.9%, 72.4%, and 89.8%, respectively. The improved neural network and multi-channel feature fusion input method can automatically detect the intestinal endometriosis area, and the detection has high robustness and accuracy, which can be used as a reference to assist doctors in clinical decision-making and intervention.