AI-powered brain microstructure imaging

点击翻译按钮获取中文摘要

This fellowship develops innovative computational methods and magnetic resonance (MR) techniques to reveal new non-invasive markers of brain microstructure. The ultimate goal is to provide non-invasive tools for improved diagnostic information as powerful as invasive techniques.Nature has built one of the most extraordinary machines we could ever conceive: the brain, using just basic cellular components of neurons and glia. Understanding how these individual components are designed (cell morphology) and assembled together (tissue microstructure) is the key to understanding both brain's structure and function, and more importantly its degeneration/dysregulation in diseases. However, it is currently impossible to quantify tissue microstructure in a non-invasive way. In fact, standard methods like histology can reveal microscopic characteristics of tissue architecture but at the cost of invasive interventions like biopsies and limited coverage of the investigated tissue, undermining diagnostic power. In contrast, sensitizing the MR imaging (MRI) contrast to water diffusion process, state-of-the-art technologies based on diffusion MRI (dMRI) provides an indirect but non-invasive probe of the tissue microstructure at the micrometer scale. This makes the acquired dMRI signal sensitive to tissue features like cellular size/shape, density, etc. Unfortunately, the tissue microstructure is highly complex while the dMRI signal is quite simple, so the mapping from signal to microstructure (inverse problem) is ill-posed. This represents a major obstacle for dMRI based techniques to replace histology as harmless diagnostic tools.To overcome these limitations, the current paradigm of microstructure imaging uses mathematical models, which relate the dMRI signal to underlying tissue properties, to estimate and map those properties by fitting the models voxel-by-voxel to dMRI data. However, there are three main limitations hampering its diagnostic power: poor sensitivity to complex tissue features that cannot be described by mathematical models; a lack of specificity to different cell types, and ambiguity in the inverse problem's solution. In this fellowship I propose a three-component shift of paradigm to address these key limitations: (i) employing detailed simulation of the tissue architecture to encode the forward problem (from tissue microstructure to MR signal), (ii) modern AI to solve the inverse problem and (iii) estimate of uncertainty to quantify ambiguity and significance of the results. I will demonstrate this in the brain by simulating normal tissue environments and those representing pathologies of common neurological diseases, like Multiple Sclerosis (MS) and Alzheimer's disease (AD). This will provide higher sensitivity to novel and important microstructural features like cell soma size/density and neurites size/complexity, leading to a new generation of quantitative imaging techniques based on water diffusion. To gain unprecedented specificity to different cell types in the brain and develop a new set of imaging markers for neurological conditions, I will combine the new paradigm with metabolites' diffusion measurements using dMR spectroscopy (dMRS). Indeed, metabolites are more cell-specific molecules than water: some are found mostly in neurons, others mostly in glia. I will prototype and validate the new technologies in controlled animal models of MS and AD and eventually provide proof-of-concept application in human patients. These innovative techniques offer great promise in the decades to come for the realisation of 'virtual histology' across a wide range of medical applications.Although the fellowship focuses on neurological diseases, it also aims to initiate follow-on projects to explore other applications, like body cancer. Alternative contrast methods will extend the methods to other MR modalities beyond diffusion for complementary and additional information on healthy and diseased tissues.