Classification of Pulmonary Images By Using Generative Adversarial Networks

Using deep learning networks has made developments in the computer vision field. Due to the growth in interstitial lung diseases (ILDs), the necessity of using modern computer-aided diagnosis (CAD) is increased. Regarding the importance of this issue, many kinds of research have been done on this subject. However, the extreme similarity between lung nodules and the complexity of detecting nodules characteristics is a severe challenge to design a system that can classify lung diseases with high accuracy. This study is done in 3 steps. First, the Convolutional Neural Network (CNN) is used due to the effectiveness and higher accuracy it can provide in comparison to the previous methods. This proposed network consists of four convolutional layers with 2×2 kernels and LeakyReLU, four average pooling layers, and three fully-connected layers. The last layer has five outputs equivalent to the considered classes: healthy, ground-glass opacity (GGO), micromodels, consolidation, reticulation. The main challenge in the field of medical images is the lack of labeled samples. So in the second step, Generative Adversarial Network (GAN) is used to generate data and increase the accuracy of the convolutional network structure by creating non-realistic but useful data in network learning. GAN structure is based on two neural networks. The generator, generates new data instances, while the discriminator, evaluates them for authenticity. The CNN which is used in the structure of discriminator can carefully detect the similarity of produced data and lung nodule classes. To train the CNN, in the third step, Interstitial Lung Diseases (ILDs) dataset (containing 3527 images) and 3200 GAN produced images is used. In the test phase for making an evaluation, real images that are extracted from the dataset are used. The accuracy of categorizing five types of lung nodules in the designed system is 88 %, which is 5 percent more than previous studies.