Identification of Lung Nodules in Computed Tomography Lung Images Using Artificial Neural Networks

Abstract

This research work involves segmentation of Lung images and identification of nodules in the Lung images. Segmentation is necessary to separate the lungs from the background image. Identification of nodules by using segmented image and the ground truth image is performed. Computer Tomography (CT) image slices from the Lung Image Database Consortium (LIDC) are used for evaluating the implemented algorithms. The three algorithms implemented for segmentation and lung nodule identification are as follows: 1. Deep learning back-propagation algorithm (BPA), 2. Deep learning BPA with radial basis function (RBF), and, 3. Quantized Deep learning BPA + CPN The BPA/ RBF/CPN are algorithms that use single hidden layer. Due to this their learning capacity of given sets of patterns is minimal. In order to increase the learning capacity of these three algorithms, the concept of deep learning is introduced. In the case of the implemented artificial neural network algorithms, the number of hidden layers is increased from one hidden layer to many hidden layers. As the number of hidden layers is increase, the number of weight matrices increase. More number of weight matrices are responsible for stored detailed information about the lung images and lung nodule details and their locations in the images. The inclusion of quantization helps in partitioning the pattern values and hence easily categorizing the features present in the patterns used for training and evaluating the algorithms. newline