SYNTHESIS OF BIOMEDICAL IMAGES BASED ON GENERATIVE ADVERSARIAL NETWORKS
Oleh Berezsky, Petro B. Liashchynskyi, Pavlo B. Liashchynskyi, A. R. Sukhovych, Taras Dolynyuk · Ukrainian Journal of Information Technology · 2019
Modern databases of biomedical images have been investigated. Biomedical imaging has been shown to be expensive and time consuming. A database of images of precancerous and cancerous breasts "BPCI2100" was developed. The database consists of 2,100 image files and a MySQL database of medical research information (patient information and image features). Generative adversarial networks (GAN) have been found to be an effective means of image generation. The architecture of the generative adversarial network consisting of a generator and a discriminator has been developed.The discriminator is a deep convolutional neural network with color images of 128×128 pixels. This network consists of six convolutional layers with a window size of 5×5 pixels. Leaky ReLU type activation function for convolutional layers is used. The last layer used a sigmoid activation function. The generator is a neural network consisting of a fully connected layer and seven deconvolution layers with a 5×5 pixel window size. Leaky ReLU activation function is used for all layers. The last layer uses the hyperbolic tangent activation function. Google Cloud Compute Instance tools have been used to train the the generative adversarial network. Generation of histological and cytological images on the basis of the generative adversarial network is conducted. As a result, the training sample for classifiers has been significantly increased.