基于深度学习的继发性肺结核CT辅助诊断模型构建及验证

Construction and evaluation of a CT-based deep learning model for the auxiliary diagnosis of secondary tuberculosis

目的:评价基于深度学习的继发性肺结核CT辅助诊断模型在临床应用中的价值。方法:回顾性收集2018年12月至2023年4月在重庆市公共卫生医疗救治中心接受胸部CT平扫的2004例患者的病例资料,分为肺部正常组(544例)、普通肺部感染组(526组)和继发性肺结核组(934例)。按照随机分组(通过R语言的sample函数实现训练集和测试集的完全随机分组)的方式,将数据集划分为训练集(1402例,70.0%)和测试集(602例,30.0%)。所有图像采用肺野自动分割算法,获得肺野区域。进一步采用BasicNet和DenseNet算法进行三组间的分类研究。采用曲线下面积(area under curve,AUC)、敏感度、特异度和准确率评价模型的分类性能。最后,在测试数据中,将最优模型与3位不同年资的放射科医生的诊断结果进行比较。结果:602例独立测试集中,DenseNet模型的性能优于BasicNet模型,两种模型的平均AUC、敏感度、特异度和准确率分别为92.1%和89.4%、79.7%和74.0%、89.4%和86.6%、86.2%和83.3%。其中,DenseNet模型的诊断性能优于低年资医生(准确率分别为90.7%和89.1%,Kappa=0.677),与中年资和高年资医生的诊断水平(准确率分别为90.7%、92.2%和95.3%,Kappa值分别为0.746和0.819)保持高度一致性。结论:DenseNet模型能较准确地识别继发性肺结核,与放射科中年资医师的诊断水准相当,可以作为继发性肺结核的辅助诊断工具。

Objective:To develop a deep learning-based auxiliary diagnostic model for secondary tuberculosis using CT scans and evaluate its clinical applicability.Methods:A retrospective collection was conducted on clinical data of 2004 patients who underwent chest CT scans at the Chongqing Public Health Medical Center from December 2018 to April 2023.The patients were divided into three groups:secondary tuberculosis(934 patients),ordinary lung infection(526 patients),and normal lungs(544 patients).Using a completely random sampling method,the dataset was divided into a training set(1402 patients,70.0%)and a test set(602 patients,30.0%).An automatic lung field segmentation algorithm was applied to isolate the lung field in all images.BasicNet and DenseNet classification algorithms were used for categorize the three groups.The discriminative performance of the model was evaluated using metrics such as area under curve(AUC),sensitivity,specificity,and accuracy.Finally,the optimal model was compared with three radiologists of different years of experience using testing data.Results:Using 602 samples in an independent test set,the DenseNet model demonstrated superior performance compared to the BasicNet model.They achieved an average AUC,sensitivity,specificity,and accuracy of 92.1%vs.89.4%,79.7%vs.74.0%,89.4%vs.86.6%,and 86.2%vs.83.3%,respectively.The diagnostic performance of the DenseNet model was superior to that of young doctors(accuracy:90.7% and 89.1%,Kappa=0.677)and exhibited high diagnostic consistency with middle and highly experienced radiologists without any significant difference(accuracy:90.7%,92.2% and 95.3%,Kappa=0.746,0.819).Conclusion:The DenseNet model can accurately identify secondary tuberculosis,achieving a competency level similar to a middle experienced radiologist,making it a potential auxiliary diagnostic tool for secondary tuberculosis.