Applications of Artificial Intelligence in Quantitative Magnetic Resonance Imaging

Abstract

Methods of quantitative magnetic resonance provide measurable insights into tissue properties and their pathophysiological changes, and therefore find applications in many clinical fields. However, the analysis of these quantitative images is extremely time consuming as well as difficult to implement, therefore these methods are currently used primarily by scientists and researchers. To address the challenges associated with implementing quantitative magnetic resonance imaging (qMRI) into routine practice, this work focuses on advanced analysis and the application of artificial intelligence (AI) approaches. A primary objective of this thesis is to automate the extraction of quantitative parameters from selected regions of interest in order to facilitate computer-aided diagnostics. In the future, this may help evaluate the quantitative properties of tissues in physiological and pathophysiological conditions, as well as monitor the progression of diseases. As a secondary goal, the thesis focuses on medical image processing, specifically the preprocessing, segmentation, and analysis of qMRI images without using artificial intelligence.

The thesis introduction includes an extensive summary of quantitative magnetic resonance methods and artificial intelligence for medical image processing, and a literature review focused on two selected topics. The first topic of interest in this thesis is the use of quantitative methods and artificial intelligence to quantify fat infiltration in the pancreas (Study I). In this study, a unique dataset with a quantitative parameter of fat fraction utilizing the Dixon technique was created, and a detailed analysis of three selected architectures was performed: U-Net, ResU-Net, and nnU-Net. A second topic focuses on Achilles tendinopathy and its quantification through T2* relaxation times (Study II). A convolutional neural network was implemented in this study to automate quantifiable changes in the tendon as well as a software application that enabled manual analysis.

The results of this work indicate the reliable use of artificial intelligence facilitating demanding quantitative analysis, paving the way for the implementation of these potential imaging biomarkers in clinical practice as well as for reducing variability. Thus, the work reflects the current trend of Open Science.