胸部X光片(CXR)是用于检测导致肺部混浊的呼吸系统疾病的最广泛可用的射线照相成像方式。CXR报告通常使用非标准化语言,导致主观、定性,和不可重现的不透明度估计。我们的目标是开发一个强大的深度迁移学习框架,并对其进行调整,以估计CXR的肺部不透明程度。根据排除标准选择CXR数据后,分割方案用于ROI(感兴趣区域)提取,以及分割的所有组合,数据平衡,和分类方法进行了测试,以选择表现最好的模型。使用多重交叉验证从初始选择的顶部模型中确定最佳模型,基于适当的性能指标,以及新颖的宏平均热图一致性评分(MAHCS)。将最佳模型的性能与专家医师注释者的性能进行比较,并制作了热图。最后,对目标患者人群进行了模型性能敏感性分析.所提出的框架适用于使用序数多类分类估计CXR肺不透明程度的特定用例。在2020年3月24日至2020年5月22日之间获得,使用了17,418名患者的38,365名前瞻性注释的CXR。我们测试了三种神经网络架构(ResNet-50、VGG-16和ChexNet),三种分割方案(无分割,肺分割,和基于脊柱检测的横向分割),和三种数据平衡策略(欠采样,两级采样,和合成少数过采样)使用38,079张CXR图像进行训练,并验证了286张图像作为经过放射科专家裁决的开箱即用数据集。根据这些实验的结果,对于肺部混浊分类,建议使用具有欠采样和无ROI分割的ResNet-50模型,基于MAE指标和HCS(热图一致性评分)的最优值。在性能度量方面,由该模型预测的不透明度分数相对于两组放射科医师分数(OR或原始读取器和OOBTR或开箱即用读取器)之间的一致性程度优于放射科医师间的不透明度分数一致性。
Chest radiographs (CXRs) are the most widely available radiographic imaging modality used to detect respiratory diseases that result in lung opacities. CXR reports often use non-standardized language that result in subjective, qualitative, and non-reproducible opacity estimates. Our goal was to develop a robust deep transfer learning framework and adapt it to estimate the degree of lung opacity from CXRs. Following CXR data selection based on exclusion criteria, segmentation schemes were used for ROI (Region Of Interest) extraction, and all combinations of segmentation, data balancing, and classification methods were tested to pick the top performing models. Multifold cross validation was used to determine the best model from the initial selected top models, based on appropriate performance metrics, as well as a novel Macro-Averaged Heatmap Concordance Score (MA HCS). Performance of the best model is compared against that of expert physician annotators, and heatmaps were produced. Finally, model performance sensitivity analysis across patient populations of interest was performed. The proposed framework was adapted to the specific use case of estimation of degree of CXR lung opacity using ordinal multiclass classification. Acquired between March 24, 2020, and May 22, 2020, 38,365 prospectively annotated CXRs from 17,418 patients were used. We tested three neural network architectures (ResNet-50, VGG-16, and ChexNet), three segmentation schemes (no segmentation, lung segmentation, and lateral segmentation based on spine detection), and three data balancing strategies (undersampling, double-stage sampling, and synthetic minority oversampling) using 38,079 CXR images for training, and validation with 286 images as the out-of-the-box dataset that underwent expert radiologist adjudication. Based on the results of these experiments, the ResNet-50 model with undersampling and no ROI segmentation is recommended for lung opacity classification, based on optimal values for the MAE metric and HCS (Heatmap Concordance Score). The degree of agreement between the opacity scores predicted by this model with respect to the two sets of radiologist scores (OR or Original Reader and OOBTR or Out Of Box Reader) in terms of performance metrics is superior to the inter-radiologist opacity score agreement.