基于深度学习在弥漫性囊腔性肺疾病分类中的应用

Study on diffuse cystic lung disease based on deep learning

目的拟基于深度学习方法,初步对几种弥漫性囊腔性肺疾病建立辅助诊断模型,同时评估各诊断模型的分类表现能力,并从中择出最优化的诊断模型,以服务临床诊断。方法收集2010年1月至2022年10月经病理证实确诊为特发性肺纤维化(IPF)、肺淋巴管平滑肌瘤病(PLAM)以及肺朗格汉斯细胞组织细胞增生症(PLCH)的患者288例(IPF 76例、PLAM 179例、PLCH 33例),CT数共877例(IPF 232例、PLAM 557例、PLCH 88例)。以2019年12月31日为节点,将CT分为数据集A(CT数共500例,其中IPF 185例、PLAM 265例、PLCH 50例)和数据集B(CT数共377例,其中IPF 47例、PLAM 292例、PLCH 38例)。将数据集A按7∶1∶2的比例随机划分为训练集、验证集、测试集,进行预处理及数据扩增后导入6种不同的深度学习神经元网络进行训练。绘制受试者工作特征曲线,以曲线下面积、准确度、敏感度、特异度、F1分数等指标评估模型性能,从中选出最优模型。再从数据集B中各病种随机抽取30例形成测试集B,投入训练好的最优模型中进行测试,再次以相同指标评估模型性能。结果基于数据集A测试结果,建立的6个诊断模型在识别分类IPF、LAM上的各项目分类性能优先,AUC皆大于0.9。EfficientNet在分类LCH上,0.6<AUC<0.7,分类效能一般;Vgg11在分类LCH上,0.8<AUC<0.9,其余4个模型在分类LCH上,AUC皆大于0.9,分类效果优秀;除InceptionV3外,其余5个诊断模型在识别分类LCH上的性能欠佳。综合多个指标考虑,InceptionV3模型在6个模型中综合性能最佳,各项评估参数都处于较高水平,总体准确率94.90%,精确率93.49%,召回率90.84%,特异度96.91%。将数据集B投入训练好的InceptionV3模型进行测试,根据输出结果计算得准确率81.11%,精确率82.50%,召回率81.11%,特异度为90.6%。结论基于胸部CT图像结合深度学习技术构建的6种识别分类模型能较有效地区分LAM、LCH、IPF三种疾病,特别是在IPF、LAM的识别分类上,其中以InceptionV3神经元网络构建的模型分类的效能最佳。

Objective To develop deep learning-based auxiliary diagnostic models for diverse pulmonary diffuse cystic diseases,and subsequently evaluate their classification performance to identify the optimal model for clinical diagnosis.Methods A total of 288 patients diagnosed with idiopathic pulmonary fibrosis(IPF),pulmonary lymphangioleiomyomatosis(PLAM),and pulmonary Langerhans cell histiocytosis(PLCH)were prospectively enrolled from the First Affiliated Hospital of Guangzhou Medical University between January 2010 and October 2022,comprising 76 cases of IPF,179 cases of PLAM,and 33 cases of PLCH.A total of 877 CT cases were collected,comprising 232 cases of IPF,557 cases of PLAM,and 88 cases of pulmonary PLCH.Based on the cutoff date of December 31,2019,the CT scans were divided into two datasets:dataset A consisted of 500 CT scans including 185 IPF cases,265 PLAM cases,and 50 PLCH cases;while dataset B comprised 377 CT scans with a distribution 19of 47 IPFcases,292 PLAMcases,and 38 PLCH cases.The Dataset A was randomly partitioned into training set,validation set,and test set in a ratio of 7∶1∶2.Subsequently,six distinct deep learning neural networks were employed for training after preprocessing and data augmentation.Receiver operating characteristic curves were generated to assess the model performance using metrics such as area under the curve(AUC),accuracy,sensitivity,specificity,and F1 score in order to identify the optimal model.Furthermore,a test set B comprising 30 randomly selected cases from dataset B for each disease type was utilized to evaluate the trained optimal model by employing the same aforementioned metrics.Results In test A,six well-established diagnostic models demonstrated superior classification performance for IPF and LAM,with an AUC greater than 0.9.For LCH,EfficientNet exhibited low classification efficiency with an AUC between 0.6 and 0.7,while Vgg11 showed an AUC between 0.8 and 0.9;the other four models displayed excellent classification efficiency with an AUC greater than 0.9.Except for Inception V3,the remaining five diagnostic models performed poorly in identifying and classifying LCH lesions.Considering multiple indicators,the InceptionV3 model showcased optimal comprehensive performance among the six models,achieving high evaluation parameters such as overall accuracy(94.90%),precision(93.49%),recall(90.84%),and specificity(96.91%).TestB was conducted using the trained InceptionV3 model resulting in an accuracy of 81%,precision of 82%,recall of 81%,and specificity of 90%.Conclusions Six recognition and classification models,developed using deep learning technology in conjunction with pulmonary CT images,demonstrate effective discrimination between LAM,LCH,and IPF.Notably,the model constructed utilizing the InceptionV3 neural network exhibits superior efficiency in accurately recognizing and classifying IPF and LAM.