Aumento de imágenes de mama mediante Redes Generativas Antagónicas (GANs)

Abstract

Abstract: Training deep neural networks requires large amounts of information to be able to specify the classification of the data, especially in a diagnosis where the data are sparse and unbalanced. Therefore, in this work thesis two algorithms based on Antagonistic Neural Networks (GAN) were implemented for the generation of synthetic images of mammography and breast ultrasound, (i) Spectral Normalization GAN (SNGAN) and (ii) Conditional GAN ​​(CGAN). To measure the similarity of the synthetic results with respect to the real images, the FID metric was used, the CGAN I present more realistic results for the mammography data set with an FID of 52.89, for On the other hand, for the ultrasound data set the best algorithm to generate images synthetic was the SNGAN with a FID of 116.85. In conclusion, despite the use of the same hyperparameters for the training of the two networks, the selection of the best algorithm GAN for data growth will depend on other factors such as the network architecture or the amount of data.