Resumo
Lung cancer corresponds to 26% of all deaths due to cancer in 2017, accounting for more than 1.5 million deaths globally. Considering this challenging situation, several computeraided diagnosis(CAD) systems have been developed to detect lung cancer at early stages, which increases the patients’ survival rate. Motivated by the success of deep learning in natural and medical image classification tasks, the main objective of this work is to explore the performance of deep transfer learning from non-medical images on the lung nodules malignancy classification task. More specifically, this task is represented in this work by a binary classification problem using the Lung Image Database Consortium and Image Database Resource Initiative (LIDC/IDRI) dataset. To accomplish that, the methodology applied in this work is divided into 4 stages. First, the LIDC/IDRI has its 1018 chest Computed Tomography (CT) exams and medical annotations processed. Then, several Convolutional Neural Networks (CNN), such as VGG16, VGG19, MobileNet, Xception, InceptionV3, ResNet50, Inception-ResNet-V2, DenseNet169, DenseNet201, NASNetMobile and NASNetLarge models, are built, trained on ImageNet dataset, converted into feature extractors and applied on the LIDC/IDRI nodule images. Second, each set of deep features is submitted to 10-Fold Cross Validations with Naive Bayes, Multilayer Perceptron (MLP), Support Vector Machine (SVM), K-Nearest Neighbors (KNN) and Random Forest (RF) classifiers. Third, each Cross Validation average result has its evaluation metrics Accuracy (ACC), Area Under the Curve (AUC), True Positive Rate (TPR), Precision (PPV), and F1-Score computed and compared. Thus, according to the results,the Deep Feature Extractor based on ResNet50 model combined with a SVM classifier (using a Radial Basis Function (RBF) kernel), achieve an AUC metric of 93.19% (the highest value not only among the evaluated combinations, but also among the related works evaluated), a TPR of 85.38%, an ACC of 88.41%, a PPV of 73.48% and a F1-Score of 78.83%. Based on these results, deep transfer learning from non-medical images proves to be a relevant strategy to extract representative features for lung nodule malignancy classification in chest CT images. Overall, the combination of ResNet50 and SVM-RBF achieve satisfactory results and is eligible to be integrated into a CAD system.