Automatic recognition of lung tumor using convolutional neural networks

Abstract

The unchecked proliferation of abnormal cells in the lung is known as lung cancer, and it is one of the major causes of death globally. The most accurate way for finding malignant lung nodules is a thorax Computed Tomography (CT) scan. A spherical lesion called a lung nodule can either be malignant or not. Lung cancer appears as rounded, white shadow nodules on the CT scan. The candidate ROIs are calculated using existing method and some blood vessels are removed using rule-based methods based on the candidate ROIs' shape features. Next, the candidate ROIs' remaining grey and texture features are calculated, and are given to the classifier to categorize the candidates. The rule-based technique has no omissions, according to experimental results, however the misclassification probability is too high. Therefore, in the proposed method, by computing the texture features from the Grey Level Co-occurrence Matrix (GLCM) in the wavelet domain, the nodules were identified and CT images were categorized using Convolutional Neural Networks (CNN) into two groups: those with and those without malignant lung nodules. In this work, two classifiers Support Vector Machine (SVM) and CNN are used for the recognized lung cancer image to determine the severity of the condition. Comparisons between SVM and CNN classifiers are done with regard to quality parameters including accuracy and sensitivity.