Lung Tumor Detection and Recognition Using Deep Convolutional Neural Networks

Abstract

Cancer with the greatest fatality rate, lung cancer, requires a biopsy to define its type to select the best course of treatment. For decades scientists have been working on a method to detect tumors earlier than the current procedures. This research proposes an approach using artificial intelligence to predict a tumor’s presence and malignancy using only computed tomography (CT) scans. This is by using the techniques of deep learning, transfer learning, and ensemble learning. This research made use of two different datasets, one with only benign, malignant, and normal cases called the IQ-OTH/NCCD dataset and another dataset with 3 carcinoma classes and a normal class. Applying convolutional neural networks (CNN)s architecture in VGG-16 using only the dataset with 3 classes, this experiment achieved a very satisfying result of 89% test accuracy with relatively low computational power. Then applying ensemble learning techniques in a VGG-16, ResNet50, InceptionV3, and EfficientNetB7 majority voting architecture to classify not only CT scans being malignant, benign, or normal, but also to classify different types of lung carcinomas, this architecture outperformed every other trial and achieved a 92.8% test accuracy.