基于深度学习方法的胸片异物检测

Detecting foreign objects in chest radiographs based on a deep learning method

为了提高胸片异物自动检测的能力,采用深度学习网络高效提取各种尺度和形状的异物影像特征,实现胸片中多种异物的自动稳定检测。在网络构建过程中,根据异物的影像特征,改进YOLO v4目标检测网络,通过在特征提取网络CSPDarkNet53中加入SE-block(Squeeze and Excitation),使模型能够区别利用各个通道的信息。实验结果表明,改进的深度学习网络在异物检测中能够实现92%的精确率和83%的召回率。因此,新的深度学习方法可用于胸片异物检测等应用场景,客观评判摄影质量,为放射影像的质量控制打下基础。

In order to improve the ability of automatically detecting foreign objects in chest radiographs,a deep learning network is used to capture the imaging features of foreign objects with various scales and shapes efficiently,thereby realizing the automatic and stable detection of various foreign objects in chest radiographs.During network construction,according to the imaging features of foreign objects,YOLO v4 network is improved by embedding SE-block(Squeeze and Excitation)into the feature extraction network CSPDarkNet53 to make the model capable of distinguishing and utilizing the information of each channel.Experimental results demonstrate that the proposed deep learning network achieves 92%accuracy and 83%recall rate for foreign object detection.Therefore,the proposed deep learning approach can be used to identify foreign objects in chest radiographs and assess image quality automatically and objectively,which lays a foundation for the quality control of radiographic images.