Magnetic Resonance imaging is a really important imaging technique distinguished by
the lack of ionizing radiation, great soft-tissue contrast and remarkable resolution. The
main drawback of performing a Magnetic Resonance Imaging (MRI) scan is time
consumption. As a result, many MRI scans are performed in a shorter time sacrificing
overall image quality. Thus, there is a need for algorithms able to convert low-resolution
images, obtained from faster scans, to super-resolution images comparable ...
Magnetic Resonance imaging is a really important imaging technique distinguished by
the lack of ionizing radiation, great soft-tissue contrast and remarkable resolution. The
main drawback of performing a Magnetic Resonance Imaging (MRI) scan is time
consumption. As a result, many MRI scans are performed in a shorter time sacrificing
overall image quality. Thus, there is a need for algorithms able to convert low-resolution
images, obtained from faster scans, to super-resolution images comparable to full-time
scans.
Generative Adversarial Networks (GANs) have the ability to generate new content,
mainly images and data distributions. This kind of network can be divided into 2 units
that act in competition, a generator and a discriminator. The generator receives an input
and must return, an output which follows the target distribution. The discriminator acts
as the judge of the process, determining which image is real and which was created by
the generator.
In this bachelor thesis, a two-stage training process is developed using the ESRGAN
framework. First, a Peak Signal to Noise Ratio (PSNR) oriented model is developed
without adversarial training, performing similar to standard deep networks. In the second
stage, the discriminator is introduced and trained together with the PSNR-oriented
generator. The second stage allows the model to learn realistic texture representations,
measured with the perceptual index (P.I.) metric.
The obtained results demonstrate the potential application of these networks to medical
imaging. The reconstructed MRI images were objectively more similar to the original
High Resolution (HR) ground truth when compared to other conventional image
reconstruction methods. The PSNR trained model yielded better PSNR and structural
similarity (SSIM) performance but with a lower perceptual index. The GAN model
showed greater P.I. by being able to replicate image textures and creating overall more
visually appealing images.
+