Using very high magnetic fields significantly improves the spatial resolution of magnetic resonance imaging data. Still, it does not reach the microscopic scale that allows visualization of individual cells. Researchers at UNIRS (NeuroSpin) and their collaborators at the Institute of NeuroScience and Medicine of Juelich are developing an original approach that relies on large-scale numerical simulation to drive machine learning algorithms to decode from simple acquisitions of diffusion MRI either from healthy volunteers or patients the cellular organization of brain tissues. Diffusion MRI is sensitive to the microscopic movement of water within tissues. This tool, based on the joint use of MRI signal simulations and artificial intelligence (AI) methods, aims to address the need for fundamental research on the human brain, in particular the in vivo decoding of cerebral cortex cytoarchitecture. It will also be helpful for clinicians by providing them with a substitute to biopsy that remains a very invasive surgical procedure.
The decoding tool is based on three simulation steps that require the use of High Performance Computing (HPC) infrastructures. A collection of ultra-realistic virtual samples of the brain tissue must be generated, and for each of them simulate the diffusion of the water and then the diffusion MRI signal that would be obtained.
In the article published in NeuroImage, the team details the principle of the new algorithm developed to simulate ultra-realistic virtual samples. The researchers demonstrate that MEDUSA algortihm (Microstructure Environment Designer with Unified Sphere Atoms) can synthesize the membrane geometry of cell populations corresponding to any brain region. Based on the prescription of a set of dozens of parameters, the algorithm allows to create virtual tissue samples with a degree of realism so far never achieved in silico.
This new tool makes it possible to simulate the expected MRI signal for any configuration of cell populations capable of fill the brain tissue, and thus to learn the signature MRI specific to each of these configurations by driving a machine learning method. Once trained, the method can then be used to decode the cellular organization from a reduced set of MRI signals. Thus it constitutes a real tool for performing virtual biopsies.
MEDUSA tool is already the subject of the first simulation campaigns within the Very Large Computing Center of the CEA (CCRT, Bruyères-le-Châtel), to demonstrate that it is possible to quantify precisely the swelling of the axons that occurs during a stroke. A larger campaign to simulate around ten billion virtual samples is currently undergoing a request for resources to GENCI
1 to enable in vivo mapping of the microstructure of the brain white matter.
This article is adapted and translated from the original article published on Scoop.it by LifeSciencesUPSaclay : http://sco.lt/9M612W
[1] Grand équipement national de calcul intensif