基于特征融合的U-Net肺自动分割方法

目的:将肺部颜色特征与纹理特征融合形成一种更有效的特征，并利用改进的U-Net深度学习网络结构对肺部CT影像进行图像分割以准确提取肺实质区域。方法:使用的CT影像数据来源于LIDC-IDRI数据库，首先通过色彩空间转换、高阶邻域统计的方法分别提取颜色特征和纹理特征，然后采用加权平均直方图融合两类特征，最后将特征输入改进后的U-Net模型，进行1 000次CT扫描测试，以达到完整的肺实质输出。结果:该方法最终的骰子系数、灵敏度、特异性分别为93%、96%和97%。结论:本方法较单一特征分割方法具有较高的分割精度，有效提高肺实质的分割精度，可为后续的肺部疾病自动诊断提供可靠基础，减少临床诊断的成本并节省医生诊断时间。

U-Net automatic lung segmentation based on feature fusion

Abstract: Objective To obtain a type of more effective feature by combining lung color features with texture features, and to accurately extract lung parenchyma by segmenting the lung CT image using improved U-Net deep learning network structure. Methods The CT image data used in the study were derived from LIDC-IDRI dataset. The color features and texture features were firstly extracted through color space conversion and high-order neighborhood statistics. Then, the mean weighted histogram was used to fuse the two types of features and the obtained features were input into the improved U-Net model for 1 000 CT scan tests, thereby achieving a complete lung parenchyma output. Results The Dice coefficient, sensitivity and specificity of the proposed method were 93%, 96% and 97%, respectively. Conclusion The proposed method which has a higher segmentation accuracy than the single feature segmentation method can effectively improve the accuracy of lung parenchyma segmentation and provide a reliable basis for the subsequent automatic diagnosis of lung diseases, thus reducing the cost of clinical diagnosis and shortening the time for diagnosis.