Abstract
Magnetic resonance imaging (MRI) scans tend to be long procedures, making it difficult for patients to remain motionless. Generative adversarial networks (GANs) are artificial intelligence models widely used in image processing. This paper addresses the use of GANs to aid in the reconstruction of magnetic resonance images made with undersampled single-coil data, aiming to reduce MRI scan time. The GAN improves quality metrics of images reconstructed with different trajectories (vertical Cartesian, random Cartesian, radial, spiral 1, and spiral 4), which present distortions due to not meeting the Nyquist–Shannon sampling theorem. Tests are performed by feeding the trained model with images made from undersampled data to generate enhanced images. The generator receives these images and attempts to produce enhanced versions based on typical MRI scans, while the discriminator compares generated and typical images to determine which are real or fake. At each epoch, both models refine themselves based on the discriminator’s verdict, improving image quality. Metrics such as the structural similarity index and signal-to-noise ratio are extracted from the produced images. In the best tests, increases of 4× in average SSIM and up to 14× in average SNR were achieved.