Deep Learning: Convolutional Neural Networks for lung cancer detection and classification

Abstract

Lung cancer is a disease caused by uncontrollable tissue growth in the lungs. This kind of malignant tumor has become the cancer with the highest mortality rate worldwide. Fortunately, when lung cancer is detected in its early stages, the survival rate of patients has great improvement. Due to the Computerized Tomography (CT) scans technology, it is possible to obtain detailed internal images of the body, and this information can be employed in the training of systems to detect and classify malignant nodules. Nevertheless, CT images normally have high dimensionality, and they require a great time effort of a specialist to be analyzed. Furthermore, these professionals need to take care of diverse patients in the meantime. Consequently, factors such as fatigue can lead them to make mistakes. Computer-Aided Diagnosis Systems are extremely useful in these cases, they can assist the specialists by automating the detection of suspected nodules, and hence improve the efficiency and quality of diagnosis. For instance, this type of system uses Machine Learning techniques to create a correlation between different features of a given disease, and eventually, learn to give accurate diagnoses. Moreover, the visual recognition process is made by Convolutional Neural Networks, which are considered state-of-the-art in the object and pattern recognition area. These models present a considerable precision gain, in exchange for their great computational costs. However, the recent improvements in both software and hardware areas play an essential role in this field, allowing the systems to build more complex Artificial Neural Networks, composed of numerous specialized layers. In particular, these systems are named Deep architectures. In conclusion, they are mathematical models that can construct complex concepts by creating a composition of simpler ones. In this work, an U-Net based architecture was employed to detect lung nodules in CT Scans. The model’s performance was evaluated against the Lung Nodule Analysis (LUNA16) Challenge results, ranking among the top 20 architectures in the challenge.