基于多标签深度网络的自动肋骨分割  

Automatic Rib Segmentation Based on Multi-label Deep Network

作　　者：王佳雯 陈胜[1] WANG Jia-wen;CHEN Sheng(School of Optical-electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《控制工程》2022年第7期1330-1336,共7页Control Engineering of China

基　　金：国家自然科学基金资助项目(81101116)。

摘　　要：为解决放射科医生在观察X射线胸片时,因肺区肋骨和软组织的相互重叠而无法准确检测病灶的问题,提出一种基于多标签系统的自动肋骨分割的深度学习模型MUS-Net。首先,将U-Net网络作为主体网络结构,采用多尺度输入层构造图像金字塔,边输出层为不同尺度层生成相应的局部预测图,多标签损失函数生成最终的肋骨分割图;然后,采用计算机辅助检测系统生成肺区掩膜,与肋骨分割图结合得到肺区肋骨分割图;最后,对实验结果进行主、客观评价并与对比模型相比较,所提模型在分割精度、视觉感知方面均更优,证明了该模型的有效性。In order to solve the problem that radiologists cannot accurately detect the lesions due to the overlapping of ribs and soft tissues in the lung region when observing X-ray chest radiographs,a deep learning model,MUS-Net,based on multi-label system for automatic rib segmentation is proposed.Firstly,the U-Net is taken as the main network structure,and the image pyramid is constructed with multi-scale input layers.The side-output layer generates the corresponding local prediction map for the layers of different scales,and the final rib segmentation map is generated by multi-label loss function.Then,the computer-aided detection system is used to produce lung masks,which is combined with the rib segmentation map to obtain the rib segmentation map of the lung region.Finally,the experimental results are evaluated subjectively and objectively,and compared with that of the contrast model.The proposed model is better in segmentation accuracy and visual perception,which proves the effectiveness of it.

关 键 词：胸片 多标签 肋骨分割 U-Net 肺区掩膜 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]