Super-Resolution of Biomedical Images with Generative Adversarial Networks and posterior Tumor Segmentation

João Luís Carrilho Guerreiro

Key information

Authors:

João Luís Carrilho Guerreiro (João Luís Carrilho Guerreiro)

Supervisors:

Nuno Ricardo da Cruz Garcia; Pedro Filipe Zeferino Aidos Tomás (Pedro Filipe Zeferino Tomás)

Published in

11/24/2022

Abstract

Magnetic Resonance Imaging (MRI) is an expensive medical imaging technique typically associated with long scanning times. MRI acquisition can be potentially accelerated by decreasing the spatial coverage and reducing the number of measured slices. However, this results in a lower MRI resolution and can eventually lead to misleading medical interpretations. An alternative solution comes from recent breakthroughs in Machine Learning, which have shown that high-resolution images can be recovered via super-resolution, particularly through Generative Adversarial Networks. This thesis conducts a review on GAN-based SR methods, exhibiting the immersive ability of GANs on upscaling MRIs by a x4 scale factor while at the same time maintaining trustworthy and high-frequency details. Despite quantitative results suggesting SRResCycGAN outperforms other popular deep learning methods in recovering x4 downgraded images, qualitative results show Beby-GAN holds the best perceptual quality and proves GAN-based methods hold the capacity to reduce medical costs and enable MRI applications where it is currently too slow or expensive. Additionally, Tumor Segmentation is utilized to validate the proficiency of GANs in the MRI reconstruction task. Tumor Segmentation of the synthesized images advocates marginal dissimilarities, thus there is a window for improvement. Furthermore, this thesis suggests that a chain of processes for a faster diagnosis can be conceived by merging both Super-Resolution and Tumor Segmentation. Essentially, tumor segmentation algorithms benefit from the improved spatial resolution derived from super-resolution. The diagnosis process is accelerated by acquiring low-resolution MRIs and subsequently upscaling them (via super-resolution) to detect tumors.