机器学习算法诊断PET/CT纵膈淋巴结性能评估

评估4种主流典型的机器学习方法（随机森林、支持向量机、AdaBoost、反向传播人工神经网络）对（¹⁸F-FDG） PET/CT影像中非小细胞肺癌纵膈淋巴结良恶性进行诊断分类的性能.先从168例病人的PET/CT影像中分割出1 397个淋巴结,对每个淋巴结提取出13种图像特征（D_short、area、volume、HU_mean（2D or 3D）、HU_contrast（2D or 3D）、SUV_mean（2D or 3D）、SUV_max（2D or 3D）、SUV_std（2D or 3D））;将提取出的13种图像特征进行组合,得到4种组合变量（“All features”、“High AUC features”、“Doctor's features”、“3D features”）;在4种组合变量下,分别从敏感性、特异性以及ROC曲线下的区域面积（AUC_ROC）3个方面对随机森林、支持向量机、AdaBoost、反向传播人工神经网络定量地进行诊断性能评估.评估结果显示,4种分类器分割结果的敏感性为77%~84%,特异性为81%~84%,AUC_ROC为0.86~0.90.在显著性（p<0.001）条件下对比发现,虽然机器学习方法的特异性略低于人类专家,但是敏感性显著优于人类专家.研究结果表明,三维图像特征及PET/CT影像组合特征可以显著提高AUC_ROC.基于上述研究结果可以得出结论,虽然4种机器学习方法在（¹⁸F-FDG） PET/CT影像的非小细胞肺癌纵膈淋巴结的良恶性诊断中展现了不错的敏感性,但它们的特异性有待进一步提高,在未来需要尝试多种分类方法进行联合实验,使用更高级的机器学习方法如深度学习进行进一步的研究.

Evaluation of machine learning classifiers for diagnosing mediastinal lymph node metastasis of lung cancer from PET/CT images

The classification performance in diagnosing mediastinal lymph node metastasis of non-small cell lung cancer (NSCLC) was evaluated from (¹⁸F-FDG) PET/CT images with four mainstream classical machine-learning classifiers (random forest, support vector machines, adaptive boosting, and back-propagation artificial neural network). 1397 lymph nodes were segmented from 168 patients' PET/CT images, and 13 kinds of image features (D_short, area, volume, HU_mean(2D or 3D), HU_contrast(2D or 3D), SUV_mean(2D or 3D), SUV_max(2D or 3D), SUV_std(2D or 3D)) were extracted from each lymph node. The extracted 13 kinds of image features were combined to get 4 kinds of combinatorial variables ("All features", "High AUC features", "Doctor's features", "3D features"). The diagnostic performance of random forest, support vector machines, adaptive boosting, and backpropagation artificial neural networks were quantitatively evaluated according to the four kinds of combinatorial variables in terms of sensitivity, specificity and area under the ROC curve (AUC_ROC). The evaluation results show that the four classifiers yielded sensitivity are between 77%-84%, specificity between 81%-84% and AUC_ROC between 0.86-0.90. Under the significant contrast conditions (p<0.001), although the specificity of machine learning methods is slightly lower than that of human experts, but the sensitivity is significantly better than that of human experts. Results showed that 3D features and PET-CT combined features resulted in significant improvement of AUC_ROC. Although the 4 kinds of machine learning methods demonstrate promising sensitivities for mediastinal lymph node metastasis of non-small cell lung cancer diagnosis from (¹⁸F-FDG) PET/CT images, their specificities still need to be improved. A variety of classification methods are needed to conduct joint experiments in the future, and more advanced machine learning methods such as deep learning will be used for the further study.