摘要
目的研究基于机器学习的影像组学方法对胸部低剂量CT(LDCT)图像中肺结节进行分类并预测随访过程中形态变化的可行性。方法建立两个回顾性队列,采集LDCT图像和肺CT筛查报告和数据系统(Lung-RADS)结构式报告中结节大小、位置、性质,随访过程中形态变化等信息,用于建模。在队列1中,首先使用深度学习算法在718个连续病例中检出肺结节,经专家清洗后得到有效直径为4~20 mm的肺结节,分为实性结节(n=317)、纯磨玻璃结节(n=185)、混合型磨玻璃结节(n=57)和钙化结节(n=128)。利用梯度提升树(GBDT)算法进行肺结节分类建模,通过交叉验证算法选择最优的GBDT分类器建模参数,然后利用最优参数建立肺结节分类模型,并验证其效能。在队列2中,仅选择初诊患者LDCT中结节分类最高为Lung-RADS 3类的病例(n=116),在12个月内再次行LDCT检查,且部分进行了>12个月的随访。根据随访结果将Lung-RADS 3类结节分为:有风险的活动性结节(n=56)和无风险的非活动性结节(n=60)。利用多变量非平衡调整逻辑回归算法(IALR)进行肺结节生长预测建模,并验证其效能。结果基于影像组学特征根据GBDT算法建立肺结节分类模型,平均ROC曲线下面积(AUC)为0.80(纯磨玻璃结节AUC=0.91,钙化结节AUC=0.90,实性结节AUC=0.81,混合型磨玻璃结节AUC=0.57)。预测Lung-RADS 3类肺结节变化的模型,其ROC曲线AUC为0.69,敏感性、特异性及准确性分别为62.3%、64.1%和62.6%。结论基于机器学习的LDCT肺结节分类有较高的准确性,对Lung-RADS 3类结节预后随访的效能仍需进一步研究。
Objective To evaluate the feasibility of machine learning methods in detection,classification and follow-up prediction of lung nodules on low dose chest CT examinations.Methods Two retrospective cohorts were constructed in this study.In the first cohort,the low dose CT images and structured reports were collected for study of lung nodule detection and classification.In the second cohort,additional follow-up information was collected for Lung-RADS 3 category nodules,for the purpose of prediction study.The following characteristics of lung nodules were collected by experienced radiologists,i.e.,location,size and classification.If follow-up images were available,the morphological changes of nodules were also recorded.Deep learning algorithms were used for detection of lung nodule,sequentially.Machine learning and radiomics algorithm were used for classification and follow-up prediction.Results 687 nodules were recruited for radiomics study.By using GBDT algorithm,the classification of different nodules was acceptable for clinical practice,with average AUC of 0.80[AUC 0.91 for pure ground-glass opacity(pGGO),AUC 0.90 for calcified nodules,AUC 0.81 for solid nodules,and AUC 0.57 for mixed ground-glass opacity(mGGO),respectively].116 nodules were recruited for follow-up prediction.By using IALR module,AUC for differentiation of active Lung-RADS 3 nodules from non-active nodules was 0.69,and the sensitivity,specificity and accuracy were 62.3%,64.1%,and 62.6%,respectively.Conclusion Radiomics methods could be used for classification of lung nodules,and it could potentially be used for follow-up prediction of Lung-RADS 3 category nodules.
作者
张晓东
邢倩
韩超
谢辉辉
刘义
王鹤
王慧慧
刘佳
毛丽
李秀丽
王霄英
ZHANG Xiaodong;XING Qian;HAN Chao(Department of Radiology,Peking University First Hospital,Beijing 100034,P.R.China)
出处
《临床放射学杂志》
CSCD
北大核心
2020年第10期1962-1966,共5页
Journal of Clinical Radiology
