Abstract
The goal of this thesis is to explore one of the use cases of AI in the medical field, that is, the possibility of generating synthetic images to aid Computer Assisted Diagnosis. Medical imaging data is often scarce and difficult to obtain for patient privacy reasons. Augmenting datasets with realistic data to train models used in diagnosis would contribute to a higher model accuracy and increase its potential of being used in the real world. This is why we explored the use of PGGAN (Progressive Growing of GANs, belonging to the class of generative adversarial networks) to generate artificial neuroimaging samples of healthy brains and brains with glioma and meningioma tumors. We trained the PGGAN model to generate, from random noise, images of progressively growing resolution (from 4x4 to 64x64). We stopped at 64x64 resolution due to time and hardware constraints, so at this level the quality and fidelity of the synthetic images could not compare to real samples, even if they contained brain-like shapes and even some sub-structures. However, we managed to confirm that using the PGGAN (doubling the image resolution in every phase) one is able to obtain images of increasing fidelity. We recognize that training a PGGAN is an assiduous process; the model is sensitive to minute changes of hyperparameters. We are convinced that employing a more powerful processing unit and giving the model more time to train, it is possible to obtain images of very high fidelity.