基于机器学习的非小细胞肺癌治疗前 F-FDG PET/CT 纵隔淋巴结诊断方法。

Machine learning-based diagnostic method of pre-therapeutic F-FDG PET/CT for evaluating mediastinal lymph nodes in non-small cell lung cancer.

机构信息

Department of Nuclear Medicine, Veterans Health Service Medical Center, Seoul, South Korea.

Department of Nuclear Medicine, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, South Korea.

出版信息

Eur Radiol. 2021 Jun;31(6):4184-4194. doi: 10.1007/s00330-020-07523-z. Epub 2020 Nov 25.

DOI:10.1007/s00330-020-07523-z

PMID:33241521

Abstract

OBJECTIVES

We aimed to find the best machine learning (ML) model using F-fluorodeoxyglucose (FDG) positron emission tomography/computed tomography (PET/CT) for evaluating metastatic mediastinal lymph nodes (MedLNs) in non-small cell lung cancer, and compare the diagnostic results with those of nuclear medicine physicians.

METHODS

A total of 1329 MedLNs were reviewed. Boosted decision tree, logistic regression, support vector machine, neural network, and decision forest models were compared. The diagnostic performance of the best ML model was compared with that of physicians. The ML method was divided into ML with quantitative variables only (MLq) and adding clinical information (MLc). We performed an analysis based on the F-FDG-avidity of the MedLNs.

RESULTS

The boosted decision tree model obtained higher sensitivity and negative predictive values but lower specificity and positive predictive values than the physicians. There was no significant difference between the accuracy of the physicians and MLq (79.8% vs. 76.8%, p = 0.067). The accuracy of MLc was significantly higher than that of the physicians (81.0% vs. 76.8%, p = 0.009). In MedLNs with low F-FDG-avidity, ML had significantly higher accuracy than the physicians (70.0% vs. 63.3%, p = 0.018).

CONCLUSION

Although there was no significant difference in accuracy between the MLq and physicians, the diagnostic performance of MLc was better than that of MLq or of the physicians. The ML method appeared to be useful for evaluating low metabolic MedLNs. Therefore, adding clinical information to the quantitative variables from F-FDG PET/CT can improve the diagnostic results of ML.

KEY POINTS

• Machine learning using two-class boosted decision tree model revealed the highest value of area under curve, and it showed higher sensitivity and negative predictive values but lower specificity and positive predictive values than nuclear medicine physicians. • The diagnostic results from machine learning method after adding clinical information to the quantitative variables improved accuracy significantly than nuclear medicine physicians. • Machine learning could improve the diagnostic significance of metastatic mediastinal lymph nodes, especially in mediastinal lymph nodes with low 18F-FDG-avidity.

摘要

目的

我们旨在找到最佳机器学习（ML）模型，使用 F-氟代脱氧葡萄糖（FDG）正电子发射断层扫描/计算机断层扫描（PET/CT）评估非小细胞肺癌的转移性纵隔淋巴结（MedLNs），并将诊断结果与核医学医师的结果进行比较。

方法

共回顾了 1329 个 MedLNs。比较了提升决策树、逻辑回归、支持向量机、神经网络和决策森林模型。比较了最佳 ML 模型的诊断性能与医师的诊断性能。将 ML 方法分为仅使用定量变量的 ML（MLq）和添加临床信息的 ML（MLc）。我们根据 MedLNs 的 FDG 摄取活性进行了分析。

结果

与医师相比，提升决策树模型获得了更高的灵敏度和阴性预测值，但特异性和阳性预测值较低。医师和 MLq 的准确性之间没有显著差异（79.8% vs. 76.8%，p=0.067）。MLc 的准确性明显高于医师（81.0% vs. 76.8%，p=0.009）。在 FDG 摄取活性低的 MedLNs 中，ML 的准确性明显高于医师（70.0% vs. 63.3%，p=0.018）。