HASLab - Other Publications

Permanent URI for this collection

http://localhost:4000/handle/123456789/186

Browse

Recent Submissions

Now showing 1 - 5 of 73
  • Item
  • Item
  • Item
  • Item
    To FID or not to FID: Applying GANs for MRI Image Generation in HPC
    ( 2024) Beatriz Cepa ; Cláudia Vanessa Brito ; António Luís Sousa ; 8840 ; 7516 ; 5638
    AbstractWith the rapid growth of Deep Learning models and neural networks, the medical data available for training – which is already significantly less than other types of data – is becoming scarce. For that purpose, Generative Adversarial Networks (GANs) have received increased attention due to their ability to synthesize new realistic images. Our preliminary work shows promising results for brain MRI images; however, there is a need to distribute the workload, which can be supported by High-Performance Computing (HPC) environments. In this paper, we generate 256×256 MRI images of the brain in a distributed setting. We obtained an FIDRadImageNetof 10.67 for the DCGAN and 23.54 for the WGAN-GP, which are consistent with results reported in several works published in this scope. This allows us to conclude that distributing the GAN generation process is a viable option to overcome the computational constraints imposed by these models and, therefore, facilitate the generation of new data for training purposes.
  • Item
    Mastering Artifact Correction in Neuroimaging Analysis: A Retrospective Approach
    ( 2024) Cláudia Vanessa Brito ; 7516
    The correction of artifacts in Magnetic Resonance Imaging (MRI) is increasingly relevant as voluntary and involuntary artifacts can hinder data acquisition. Reverting from corrupted to artifact-free images is a complex task. Deep Learning (DL) models have been employed to preserve data characteristics and to identify and correct those artifacts. We propose MOANA, a novel DL-based solution to correct artifacts in multi-contrast brain MRI scans. MOANA offers two models: the simulation and the correction models. The simulation model introduces perturbations similar to those occurring in an exam while preserving the original image as ground truth; this is required as publicly available datasets rarely have motion-corrupted images. It allows the addition of three types of artifacts with different degrees of severity. The DL-based correction model adds a fourth contrast to state-of-the-art solutions while improving the overall performance of the models. MOANA achieved the highest results in the FLAIR contrast, with a Structural Similarity Index Measure (SSIM) of 0.9803 and a Normalized Mutual Information (NMI) of 0.8030. With this, the MOANA model can correct large volumes of images in less time and adapt to different levels of artifact severity, allowing for better diagnosis.