基于深度学习模型的X线胸片去骨的初步研究  

Clinical application of deep learning model for bone suppression of X-ray chest radiographs

作　　者：毕忠旭 李子元[1] 周宇[1] 任昕 李津书[1] 刘建新[1] 王霄英[1] 张晓东[1] BI Zhong-xu;LI Zi-yuan;ZHOU Yu(Department of Radiology,Peking University First Hospital,Beijing 100034,China)

机构地区：[1]北京大学第一医院医学影像科,北京100034

出　　处：《放射学实践》2021年第12期1582-1586,共5页Radiologic Practice

基　　金：北京大学第一医院青年临床研究专项基金项目(2018CR25)

摘　　要：目的:探讨基于深度学习方法对X线胸片(CXR)进行去骨处理的可行性。方法:总样本数据集由开源数据集和自采数据集两部分组成,共300个图像对,每对图像包括一个常规CXR和一个去骨后的CXR,其中开源数据集(JSRT+BSE JSRT)241对,自采数据集(双能减影胸部X光片)59对。使用深度残差网络(ResNet)作为去骨模型的基础架构,结合动态U-Net网络,并以Imagenet预训练VGG(Visual Geometry Group)网络,提取图像激活及风格特征组成损失函数,来训练去骨模型。将总样本数据随机分为训练集(80%)、调优集(5%)和测试集(15%)。以测试集的峰值信噪比(PSNR)和结构相似性(SSIM)结果作为CXR去骨模型的评价指标。结果:在测试集中,CXR去骨模型的PSNR(dB)和SSIM分别为31.94±2.49和93.37%±5.11%。根据PSNR值分析,88.89%的模型去骨后图像的质量较好,根据SSIM值分析,86.67%的模型去骨后图像与目标图像的结构相似性达到88%以上。结论:深度学习模型对X线胸片图像的去骨处理是可行的,可用于进一步研究并尝试应用于临床。Objective:To investigate the feasibility of a deep learning model-based approach to achieve clinical application for bone suppression of the chest X-ray(CXR)images.Methods:The dataset consisted of open source dataset and self-collected dataset of 300 pairs,each pair of images including a conventional X-ray chest radiograph and a bone suppressed X-ray chest radiograph,in which the open source data group(JSRT+BSE JSRT)contained 241 pairs,and the self-collected data group(dual-energy subtraction chest radiography)contained 59 pairs.The residual deep neural network(ResNet)was used as the infrastructure for the bone suppressed model,combined with the dynamic U-Net network and the Imagenet pre-trained VGG(Visual Geometry Group)network based image activation and style features were extracted to form a loss function to train the bone suppression model.The data were randomly divided into training dataset(80%),validation dataset(5%)and testing dataset(15%).The peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)in the test set were used as the evaluation metrics for the CXR bone suppression model.Results:In the test set,the PSNR(dB)and SSIM of the CXR bone suppression model were 31.94±2.49 and 93.37%±5.11%,respectively.According to the analysis of PSNR values,88.89%of the predicted images showed good quality,and according to the analysis of SSIM values,86.67%of the predicted images have more than 88%structural similarity with the target images.Conclusion:The deep learning model is feasible for bone suppression of X-ray chest images,which not only improves the diagnostic efficiency of physicians,but also greatly increases the diagnostic accuracy,and can be used for further research and potential clinical application.

关 键 词：深度学习 人工智能 X线胸片 去骨模型 

分 类 号：R814.41[医药卫生—影像医学与核医学] R734.2[医药卫生—放射医学]