基于18F-FDG PET/CT影像组学与机器学习辅助孤立性肺病变分类决策
本研究纳入769例病理确诊的原发性或转移性肺癌患者,从半自动分割的PET和CT图像中提取影像组学特征,评估5种特征选择方法与9种分类方法组合建立的模型区分原发与转移性肺病变及组织学亚型的能力。结果显示多数分类器AUC大于0.75,PET数据集区分原发与转移最高AUC为0.983(GBDT特征选择+GBDT分类),CT数据集最高AUC为0.828(XGBoost特征选择+随机森林分类)。
要点
- PET/CT影像组学联合机器学习可区分原发性与转移性肺病变及组织学亚型。
- PET数据集区分原发与转移最高AUC为0.983(GBDT+GBDT)。
- GBDT为最优特征选择方法,PET与CT数据集最优分类器分别为GBDT和随机森林。
核心结论
- 18F-FDG PET/CT影像组学特征联合机器学习方法可区分原发性与转移性肺病变并识别肺癌组织学亚型。
- GBDT和随机森林分别被认为是PET和CT数据集的最优分类方法,GBDT为本分析中的最优特征选择方法。
研究设计与样本
- 本研究纳入769例病理确诊为原发性或转移性肺癌的患者。
- 使用LIFEx软件包从同一感兴趣体积内半自动分割的PET和CT图像中提取影像组学特征。
- 患者被随机分配至训练集和验证集,通过评估5种特征选择方法和9种分类方法建立判别模型,采用十折交叉验证控制流程稳健性,以受试者工作特征曲线下面积(AUC)评价模型性能。
主要发现
- 区分原发性与转移性肺病变时,PET数据集中GBDT特征选择联合GBDT分类器获得最高AUC为0.983,CT数据集中XGBoost特征选择联合随机森林分类器获得最高AUC为0.828。
- 区分鳞状细胞癌与腺癌时,PET数据集中GBDT特征选择联合GBDT分类获得最高AUC为0.897,CT数据集中GBDT特征选择联合随机森林分类获得最高AUC为0.839。
- 多数基于决策树(DT)的模型存在过拟合,提示该分类方法不适合实际应用。
关键数值与推荐
- PET数据集区分原发与转移性肺病变最高AUC为0.983(GBDT特征选择+GBDT分类)。
- CT数据集区分原发与转移性肺病变最高AUC为0.828(XGBoost特征选择+随机森林分类)。
- PET数据集区分鳞癌与腺癌最高AUC为0.897(GBDT特征选择+GBDT分类)。
- CT数据集区分鳞癌与腺癌最高AUC为0.839(GBDT特征选择+随机森林分类)。
临床怎么用
- 对于孤立性肺病变,可利用18F-FDG PET/CT影像组学联合机器学习辅助区分原发与转移性病变及组织学亚型。
- 在PET数据集上可优先考虑GBDT特征选择联合GBDT分类,在CT数据集上可优先考虑XGBoost特征选择联合随机森林分类。
- 应避免使用基于决策树的模型进行实际分类决策,因其多存在过拟合。
局限性
- 多数基于决策树(DT)的模型存在过拟合,提示该分类方法不适合实际应用。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
PURPOSE: This study was designed and performed to assess the ability of 18F-fluorodeoxyglucose (FDG) positron emission tomography (PET) and computed tomography (CT) radiomics features combined with machine learning methods to differentiate between primary and metastatic lung lesions and to classify histological subtypes. Moreover, we identified the optimal machine learning method. METHODS: A total of 769 patients pathologically diagnosed with primary or metastatic lung cancers were enrolled. We used the LIFEx package to extract radiological features from semiautomatically segmented PET and CT images within the same volume of interest. Patients were randomly distributed in training and validation sets. Through the evaluation of five feature selection methods and nine classification methods, discriminant models were established. The robustness of the procedure was controlled by tenfold cross-validation. The model's performance was evaluated using the area under the receiver operating characteristic curve (AUC). RESULTS: Based on the radiomics features extracted from PET and CT images, forty-five discriminative models were established. Combined with appropriate feature selection methods, most classifiers showed excellent discriminative ability with AUCs greater than 0.75. In the differentiation between primary and metastatic lung lesions, the feature selection method gradient boosting decision tree (GBDT) combined with the classifier GBDT achieved the highest classification AUC of 0.983 in the PET dataset. In contrast, the feature selection method eXtreme gradient boosting combined with the classifier random forest (RF) achieved the highest AUC of 0.828 in the CT dataset. In the discrimination between squamous cell carcinoma and adenocarcinoma, the combination of GBDT feature selection method with GBDT classification had the highest AUC of 0.897 in the PET dataset. In contrast, the combination of the GBDT feature selection method with the RF classification had the highest AUC of 0.839 in the CT dataset. Most of the decision tree (DT)-based models were overfitted, suggesting that the classification method was not appropriate for practical application. CONCLUSION: 18F-FDG PET/CT radiomics features combined with machine learning methods can distinguish between primary and metastatic lung lesions and identify histological subtypes in lung cancer. GBDT and RF were considered optimal classification methods for the PET and CT datasets, respectively, and GBDT was considered the optimal feature selection method in our analysis.
引用
- Zhou Y, Ma XL, Zhang T, et al. Use of radiomics based on 18F-FDG PET/CT and machine learning methods to aid clinical decision-making in the classification of solitary pulmonary lesions: an innovative approach. Eur J Nucl Med Mol Imaging. 2021. PMID: 33547553. doi:10.1007/s00259-021-05220-7
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。