基于DR图像的多模型融合用于壹期尘肺病筛查的研究

A Study on Multi-Model Fusion Based on DR Images for the Screening of Stage I Pneumoconiosis

摘要: 目的：基于胸部X线片影像，分别构建放射组学模型与2D深度学习模型，通过多因素Logistic回归对两种模型独立预测评分进行决策级融合，构建尘肺病早期筛查联合模型，评估其临床价值，为模型选择及临床转化提供依据。材料与方法：回顾性收集193例壹期尘肺病患者及178例健康体检者胸部X线片及临床资料，采用ITK-SNAP软件在胸部X线片上完成双侧肺野感兴趣区(ROI)的手动勾画，提取高通量放射组学特征，构建放射组学模型；同时基于勾画ROI为输入构建2D深度学习模型。通过多因素Logistic回归对两种单一模型的独立预测风险评分进行决策级融合，构建联合预测模型。以受试者工作特征曲线下面积、准确率、灵敏度和特异度为核心指标，通过训练集(70%)与独立测试集(30%)分层验证模型性能。采用校准曲线、决策曲线分析DCA (Decision Curve Analysis)评估模型校准性能与临床净获益，并通过DeLong检验比较模型间的区分能力。引入三位均具有职业性尘肺病诊断资质放射科医师(高、中、低年资各1名)进行独立读片诊断，作为人工对照，采用双盲设计进行独立读片诊断。结果：在独立测试集中，放射组学模型、2D深度学习模型及联合模型的准确率依次为72.1%、66.7%、75.7%，AUC值分别为0.800 (95% CI: 0.718~0.882)、0.803 (95% CI: 0.718~0.888)、0.813 (95% CI: 0.731~0.894)。联合模型灵敏度为75.4%，特异度为75.9%，鉴别效能显著优于各单一模型。校准曲线结果显示，联合模型预测概率与实际进展的发生概率一致性良好。决策曲线分析(DCA)结果表明，在多数阈值概率范围内，联合模型的净获益均高于单一模型。DeLong检验证实，放射组学模型、2D深度学习模型与联合模型的AUC值两两比较，差异均无统计学意义。在人机对比中，联合模型的准确率及AUC值均远高于中低年资医师，略低于高年资医师。结论：联合模型在尘肺病早期DR胸片筛查中展现出良好的诊断效能与临床净获益，整体性能优于单一模型，可作为职业健康首筛的核心工具。

Abstract: Objective: Based on chest X-ray images, we constructed radiomics models and 2D deep learning models, respectively. We then performed decision-level fusion of the independent predictive scores from these two models using multivariate logistic regression to develop a combined model for the early screening of pneumoconiosis. We evaluated the clinical value of this model to provide a basis for model selection and clinical translation. Materials and Methods: We retrospectively collected chest X-ray images and clinical data from 193 patients with Stage I pneumoconiosis and 178 healthy individuals. Using ITK-SNAP software, we manually delineated regions of interest (ROIs) in both lung fields on the chest X-ray images, extracted high-throughput radiomic features, and constructed a radiomic model. Concurrently, we built a 2D deep learning model using the delineated ROIs as input. We performed decision-level fusion of the independent predictive risk scores from the two standalone models using multivariate logistic regression to construct a combined predictive model. Using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity as core metrics, we validated model performance through stratified validation using a training set (70%) and an independent test set (30%). Calibration curves and Decision Curve Analysis (DCA) were used to evaluate model calibration performance and clinical net benefit, and the DeLong test was employed to compare the discriminatory capabilities of the models. Three radiologists, each qualified to diagnose occupational pneumoconiosis (one each with high, medium, and low years of experience), were recruited to perform independent film readings as a manual control, using a double-blind design. Results: In the independent test set, the accuracy rates of the radiomics model, the 2D deep learning model, and the combined model were 72.1%, 66.7%, and 75.7%, respectively. with AUC values of 0.800 (95% CI: 0.718~0.882), 0.803 (95% CI: 0.718~0.888), and 0.813 (95% CI: 0.731~0.894), respectively. The combined model had a sensitivity of 75.4% and a specificity of 75.9%, demonstrating significantly superior diagnostic performance compared to each individual model. Calibration curve results showed good agreement between the combined model’s predicted probability and the actual probability of disease progression. Decision curve analysis (DCA) results indicated that, across most threshold probability ranges, the net benefit of the combined model was higher than that of any single model. The DeLong test confirmed that, when comparing AUC values between the radiomics model, the 2D deep learning model, and the combined model in pairs, the differences were not statistically significant. In a human-machine comparison, the accuracy and AUC values of the combined model were significantly higher than those of junior and mid-career physicians, and slightly lower than those of senior physicians. Conclusion: The combined model demonstrates good diagnostic performance and clinical net benefit in the early screening of pneumoconiosis using DR chest radiographs. Its overall performance is superior to that of single models, and it can serve as a core tool for initial screening in occupational health.