Lung nodule malignancy classification in chest computed tomography images using transfer learning and convolutional neural networks

Lung cancer accounts for more than 1.5 million deaths worldwide, and it corresponded to 26% of all deaths due to cancer in 2017. However, lung computer-aided diagnosis systems developed to identify lung cancer at early stages are increasing survival rates. This study explores the performance of deep transfer learning from non-medical images on lung nodule malignancy classification tasks in order to improve such systems. Initially, the 1018 chest computed tomography (CT) examinations and medical annotations from the LIDC/IDRI were processed. Then, several convolutional neural networks (VGG16, VGG19, MobileNet, Xception, InceptionV3, ResNet50, InceptionResNetV2, DenseNet169, DenseNet201, NASNetMobile and NASNetLarge) were built, trained on the ImageNet dataset, converted into feature extractors and applied on the LIDC/IDRI nodule images. Following this, each set of deep features was submitted to 10-fold cross-validations with naive Bayes, multilayer perceptron, support vector machine (SVM), K-nearest neighbors KNN and random forest classifiers. Finally, the evaluation metrics accuracy (ACC), area under the curve (AUC), true positive rate (TPR), precision (PPV) and F1-score of each cross-validation average result were computed and compared. The results showed that the deep feature extractor based on the ResNet50 and the SVM RBF classifier, achieved an AUC metric of 93.1% (the highest value not only among the evaluated combinations, but also among the related works in the literature evaluated), a TPR of 85.38%, an ACC of 88.41%, a PPV of 73.48% and an F1-score of 78.83%. Based on these results, deep transfer learning proves to be a relevant strategy to extract representative features from lung nodule CT images.