机器学习辅助诊断肺间质纤维化的研究

Study on the application of machine learning in auxiliary diagnosis of pulmonary interstitial fibrosis

目的:构建基于机器学习、数字图像处理技术的计算机辅助诊断系统,研究高分辨率计算机断层扫描(HRCT)技术对肺间质纤维化早筛中的应用效果。方法:选取在医院呼吸内科就诊的131例患者资料数据作为样本,按照7∶3比例将其分为训练组(91例)和测试组(40例)。基于尺度不变特征转换(SIFT)图像特征实现肺实质组织分割;采用图像配准方法将肺实质组织划分区域;应用深度学习图像检测算法检测CT图像中蜂窝影、网格影、磨玻璃影及组织增厚等诊断特征;基于CT影像检测特征结合患者问诊信息形成特发性肺纤维化(IPF)辅助诊断决策;最后与临床低年资实习医生诊断测试的准确率、测试速度、召回率、受试者工作特征曲线下面积(AUC)等结果进行对比。结果:经功能及性能测试,深度学习图像检测算法可实现CT图像预处理、CT图像肺部区域配准、CT图像中肺实质分割、蜂窝影、网格影和磨玻璃影(GGO)检测、组织增厚区域检测以及辅助诊断等功能,计算机辅助诊断方法诊断准确率可达到90.00%,与低年资实习医生比较,诊断准确率差异无统计学意义;采用机器学习辅助诊断时间平均为26.2 s,诊断时间较之放射科低年资实习医生缩短41.39%。结论:机器学习的计算机辅助诊断方法在诊断准确度方面接近放射科低年资实习医生水平,提高了图像分析效率,降低医生的工作量,未来将有助于提高IPF早期筛查效率及临床应用范围,对临床治疗具有参考意义和研究价值。

Objective:To construct a computer-aided diagnosis system through machine learning and digital image processing technique,so as to understand the application scope of high-resolution computed tomography(HRCT)technique in the early screening of pulmonary interstitial fibrosis.Methods:The data of 131 patients who admitted to respiratory department of Shunyi Hospital of Beijing hospital of traditional Chinese medicine hospital were selected as samples and they were divided into training group(91 cases)and test group(40 cases)according to the ratio of 7:3.Lung parenchyma segmentation was realized based on the image features of scale invariant feature transform(SIFT).The lung parenchyma was divided into regions by image registration method.The in-depth learning image detection algorithm was used to detect the diagnostic features such as honeycomb shadow,grid shadow,ground glass shadow and tissue thickening in CT images.The auxiliary diagnosis decision of idiopathic pulmonary fibrosis(IPF)was formed based on the detection characteristics of CT images combined with consultation information of patients.Finally,the accuracy and speed of test,recall rate and AUC of computer-aided diagnosis system were compared with the diagnosis results of clinical junior interns.Results:The results of function and performance test indicated the in-depth learning image detection algorithm could realize a series of functions included CT image preprocessing,lung region registration of CT image,lung parenchyma segmentation,honeycomb shadow,grid shadow and ground glass shadow(GGO)detection,tissue thickening region detection in CT images and auxiliary diagnosis.The diagnostic accuracy of computer-aided diagnosis method could reach 90.00%,which was not significant difference with that of junior interns by chi square test.The average time of auxiliary diagnosis with machine learning was 26.2 s,which was shortened by 41.39%than that of junior interns in radiology department.Conclusion:The computer-aided diagnosis method with machine learning is close to the level of junior interns of radiology department in diagnostic accuracy,which improves the efficiency of image analysis and reduces the workload of doctors.In future,it will help to improve the efficiency of early screening of idiopathic pulmonary fibrosis(IPF)and clinical application scope,and it will be of reference significance and research value for clinical treatment.