Optimization and analysis through deep learning for magnetic resonance imaging based on incoherent spins movement applied to brain tumors

Abstract

Magnetic resonance images (MRI) stand out for their non-invasive nature and for providing different contrast mechanisms, highlighting different characteristics of biological tissues. One of these contrast mechanisms is based on the movement of spins through both diffusion and perfusion processes. A diffusion-weighted imaging method called Intravoxel Incoherent Motion (IVIM) uses magnetic field gradients of different strengths to constrain the diffusion motion of atoms. Thus, through measured signal analysis models, it is possible to separate the intra- and extra-vascular contributions of movement, which can be useful for the understanding of various diseases, such as dementia, stroke, and tumors. This project aims to process and analyze IVIM images from healthy subjects and patients with brain tumors to determine optimized acquisition parameters and obtain more robust quantitative maps. Through Deep Learning models, we intend to create neural networks and train them with data from healthy subjects, to investigate their ability to estimate these quantitative maps for patients with brain tumors. We also aim to find the minimum data set for processing so that it is possible to point to an optimization of the acquisition protocol. Although deep learning models are proving useful for processing IVIM images, there is still no work that explores brain tumors. (AU)