| dc.description.abstract | Dementia is the seventh leading cause of death among all diseases and increases rapidly. With 10 million new cases every year, research is crucial for finding a treatment to cure dementia in the future. Magnetic resonance imaging (MRI) examination enables qualified professionals to analyze and detect discrepancies and anomalies in the brain. The quality and the signal-to-noise ratio (SNR) of MRI scans are directly proportional to the magnetic field strength used. For example, machines using a magnetic field strength of 3.0 Tesla (T) can generate scans with a higher SNR than magnetic field strengths of 1.5T and 0.5T but require more facilitation on the premises and considerable financial resources.
This thesis will explore how generative adversarial networks can improve the level of detail and SNR of MRI scans from the 1.5T domain to approach that of the 3.0T domain. GANs have proven to perform satisfactorily in similar scenarios, but only in binary classification tasks. This thesis investigates how the Pix2Pix GAN can be modified to use three-dimensional images. Furthermore, this thesis evaluates the performance through a multi-class convolutional neural network (CNN), classifying cognitively normal, mild cognitive impairment, and Alzheimer’s disease.
An average performance measure of 0.84 and an average AUC score of 0.949 was achieved by classifying the generated 3.0T* MRI images, improving the evaluation of the 1.5T domain from 0.80 and 0.710, respectively. However, the small dataset size and the short training duration of the GAN could be limitations for the GAN performance. Nevertheless, this work presents a clear potential for increasing the SNR ratio for the 1.5T domain, which could be expanded to the 0.5T domain. | |