The purpose of this research project is to observe if the GAN architectures through the generation of artificial/synthetic images, have the capacity to combine the features of different datasets (different origin) in order to enhance the accuracy metric in the classification process of medical images and improve the generalization of the classification models; for this specific case the experiments are applied on X-ray images (mammographies). As an evaluation method, the pattern in the distribution of features of the datasets will be analyzed and a comparison will be made with sets created from a synthetic image. Based on the data obtained in the different experimental phases, there is a divergent trend in the characteristics of the image sets of different origins, despite representing the same pathology and being visually similar. Furthermore, a slight decrease in the accuracy value could be observed when training classification algorithms on multiple datasets. No evidence was found to show that the use of a StyleGAN2 architecture made it possible to combine feature distributions from a common latent space; one of the possible causes to be analyzed is the effect that the low variability in the artificial images had on the feature extractors.