Classification of non-small cell lung cancer by histologic subtype using deep learning in public and private data sets of computed tomography images

Abstract Objective: To develop a deep learning system to classify non-small cell lung cancer (NSCLC) by histologic subtype—adenocarcinoma or squamous cell carcinoma (SCC)—from computed tomography (CT) images in which the tumor regions were segmented, comparing our results with those of similar studies conducted in other countries and evaluating the accuracy of automated classification by using data from the Instituto Nacional de Câncer, Brazil. Materials and Methods: To develop the classification system, we employed a 2D U-Net neural network for semantic segmentation, with data augmentation and preprocessing steps. It was pretrained on 28,506 CT images from The Cancer Image Archive, a private database, and validated on 2,015 of those images. To develop the classification algorithm, we used a VGG16-based network, modified for better performance, with 3,080 images of adenocarcinoma and SCC from the Instituto Nacional de Câncer database. Results: The algorithm achieved an accuracy of 84.5% for detecting adenocarcinoma and 89.6% for detecting SCC, with sensitivities of 91.7% and 90.4%, respectively, which are considered satisfactory when compared with the values obtained in similar studies. Conclusion: The system developed appears to provide accurate automated detection, as well as tumor segmentation and classification of NSCLC subtypes of a local population using deep learning networks trained using public image data sets. This method could assist oncological radiologists by improving the efficiency of preliminary diagnoses