A mesoscale ex vivo macaque white matter atlas using high-performance computing global tractography
Presented During:
Wednesday, June 26, 2024: 11:30 AM - 12:45 PM
COEX
Room:
Conference Room E 1
Poster No:
1600
Submission Type:
Abstract Submission
Authors:
Simon Legeay1,2, Maëlig Chauvel3, Fanny Darrault4, Guillaume Dannhoff4, Bastien Herlin1, Felix Matuschke5,2, Christophe Destrieux4,6, Markus Axer5,7,2, Ivy Uszynski1,2, Frédéric Andersson4, Igor Lima Maldonado4,6, Cyril Poupon1,2
Institutions:
1BAOBAB, NeuroSpin, Université Paris-Saclay, CNRS, CEA, Gif-Sur-Yvette, France, 2AIDAS Joint Institute, Research Centre Jülich, CEA, Jülich, Germany, Gif-sur-Yvette, France, 3Department of Neurophysics, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany, 4UMR 1253, iBrain, Université de Tours, Inserm, Tours, France, 5Institute of Neuroscience and Medicine (INM-1), Research Centre Jülich, Jülich, Germany, 6CHRU de Tours, Tours, France, 7Department of Physics, University of Wuppertal, Wuppertal, Germany
First Author:
Simon Legeay
BAOBAB, NeuroSpin, Université Paris-Saclay, CNRS, CEA|AIDAS Joint Institute, Research Centre Jülich, CEA
Gif-Sur-Yvette, France|Jülich, Germany, Gif-sur-Yvette, France
BAOBAB, NeuroSpin, Université Paris-Saclay, CNRS, CEA|AIDAS Joint Institute, Research Centre Jülich, CEA
Gif-Sur-Yvette, France|Jülich, Germany, Gif-sur-Yvette, France
Co-Author(s):
Maëlig Chauvel
Department of Neurophysics, Max Planck Institute for Human Cognitive and Brain Sciences
Leipzig, Germany
Department of Neurophysics, Max Planck Institute for Human Cognitive and Brain Sciences
Leipzig, Germany
Felix Matuschke
Institute of Neuroscience and Medicine (INM-1), Research Centre Jülich|AIDAS Joint Institute, Research Centre Jülich, CEA
Jülich, Germany|Jülich, Germany, Gif-sur-Yvette, France
Institute of Neuroscience and Medicine (INM-1), Research Centre Jülich|AIDAS Joint Institute, Research Centre Jülich, CEA
Jülich, Germany|Jülich, Germany, Gif-sur-Yvette, France
Christophe Destrieux
UMR 1253, iBrain, Université de Tours, Inserm|CHRU de Tours
Tours, France|Tours, France
UMR 1253, iBrain, Université de Tours, Inserm|CHRU de Tours
Tours, France|Tours, France
Markus Axer
Institute of Neuroscience and Medicine (INM-1), Research Centre Jülich|Department of Physics, University of Wuppertal|AIDAS Joint Institute, Research Centre Jülich, CEA
Jülich, Germany|Wuppertal, Germany|Jülich, Germany, Gif-sur-Yvette, France
Institute of Neuroscience and Medicine (INM-1), Research Centre Jülich|Department of Physics, University of Wuppertal|AIDAS Joint Institute, Research Centre Jülich, CEA
Jülich, Germany|Wuppertal, Germany|Jülich, Germany, Gif-sur-Yvette, France
Ivy Uszynski
BAOBAB, NeuroSpin, Université Paris-Saclay, CNRS, CEA|AIDAS Joint Institute, Research Centre Jülich, CEA
Gif-Sur-Yvette, France|Jülich, Germany, Gif-sur-Yvette, France
BAOBAB, NeuroSpin, Université Paris-Saclay, CNRS, CEA|AIDAS Joint Institute, Research Centre Jülich, CEA
Gif-Sur-Yvette, France|Jülich, Germany, Gif-sur-Yvette, France
Igor Lima Maldonado
UMR 1253, iBrain, Université de Tours, Inserm|CHRU de Tours
Tours, France|Tours, France
UMR 1253, iBrain, Université de Tours, Inserm|CHRU de Tours
Tours, France|Tours, France
Cyril Poupon
BAOBAB, NeuroSpin, Université Paris-Saclay, CNRS, CEA|AIDAS Joint Institute, Research Centre Jülich, CEA
Gif-Sur-Yvette, France|Jülich, Germany, Gif-sur-Yvette, France
BAOBAB, NeuroSpin, Université Paris-Saclay, CNRS, CEA|AIDAS Joint Institute, Research Centre Jülich, CEA
Gif-Sur-Yvette, France|Jülich, Germany, Gif-sur-Yvette, France
Introduction:
While global tractography offers enhanced accuracy and reliability compared to standard streamline tractography methods [1], it comes at the cost of substantial computational resources in terms of both time and memory. This approach simultaneously generates and optimizes the trajectories of virtual axonal white matter (WM) fibers (represented as connected spin-glasses), considering the local orientation distribution derived from diffusion MRI (dMRI) data. Leveraging this method on extensive datasets poses significant challenges, leading to the creation of ExaTract [2], a novel High-Performance Computing (HPC) global tractography approach.
In this study, we demonstrate the ability of ExaTract to reconstruct numerous brain connections and robustly identify deep WM bundles from a very high resolution (250μm) dataset in a reasonable time.
In this study, we demonstrate the ability of ExaTract to reconstruct numerous brain connections and robustly identify deep WM bundles from a very high resolution (250μm) dataset in a reasonable time.
Methods:
Ex vivo brain: The brain of a 3-year-old female Macaca fascicularis was collected by the iBrain Unit (University of Tours, France).
Imaging protocol: The sample was scanned using a Bruker Biospin 11.7 T MRI system equipped with a 60mm volume coil. The imaging protocol included a 3D T2-weighted (T2w) MSME 100μm scan, a 180μm 2D T2w MSME scan, and 250μm multishell dMRI scans using 3D segmented EPI PGSE sequences at b=1500/4500/8000s/mm2 along 25/60/90 directions (TE/TR=24/250ms, 33 segments) [3].
Pre-processing: After noise correction, dMRI data were used to compute an ODF map based on the analytical Q-ball model. A tractography mask was manually delineated by 2 independent neuroanatomists from the T2w MRI previously matched to the dMRI dataset using ANTs [4]. The mask was split into 160 partitions using METIS [5] to distribute global tractography optimization across supercomputer nodes while minimizing partition interfaces for communication cost preservation.
Global tractography: According to the previous partitioning, ExaTract was launched on 20 nodes of the Joliot-Curie HPC cluster, with 160 MPI[6] processes, each with 16 multithreaded CPU cores, totaling 2560 CPU cores. The tractography framework was largely inspired by the works of [7,8] and used the following parameters: 4 initial spin-glasses/voxel, spin-glass length of 120um, connection maximum aperture angle 20°, ratio of connection/motion/creation/deletion=1/4/0.8/0.05, 1 simulated annealing cycle of initial/final temperature 0.1/0.03K, energy: connection likelihood L=4, αextern/αintern=1/3.
Post-processing: A hierarchical fiber clustering algorithm [9] was applied using the Ginkgo toolbox [10] to classify fibers with close geometries into fasciculi. Deep WM bundle-contributing fasciculi were selected using identified ROIs at the crossroads of the target bundle.
Imaging protocol: The sample was scanned using a Bruker Biospin 11.7 T MRI system equipped with a 60mm volume coil. The imaging protocol included a 3D T2-weighted (T2w) MSME 100μm scan, a 180μm 2D T2w MSME scan, and 250μm multishell dMRI scans using 3D segmented EPI PGSE sequences at b=1500/4500/8000s/mm2 along 25/60/90 directions (TE/TR=24/250ms, 33 segments) [3].
Pre-processing: After noise correction, dMRI data were used to compute an ODF map based on the analytical Q-ball model. A tractography mask was manually delineated by 2 independent neuroanatomists from the T2w MRI previously matched to the dMRI dataset using ANTs [4]. The mask was split into 160 partitions using METIS [5] to distribute global tractography optimization across supercomputer nodes while minimizing partition interfaces for communication cost preservation.
Global tractography: According to the previous partitioning, ExaTract was launched on 20 nodes of the Joliot-Curie HPC cluster, with 160 MPI[6] processes, each with 16 multithreaded CPU cores, totaling 2560 CPU cores. The tractography framework was largely inspired by the works of [7,8] and used the following parameters: 4 initial spin-glasses/voxel, spin-glass length of 120um, connection maximum aperture angle 20°, ratio of connection/motion/creation/deletion=1/4/0.8/0.05, 1 simulated annealing cycle of initial/final temperature 0.1/0.03K, energy: connection likelihood L=4, αextern/αintern=1/3.
Post-processing: A hierarchical fiber clustering algorithm [9] was applied using the Ginkgo toolbox [10] to classify fibers with close geometries into fasciculi. Deep WM bundle-contributing fasciculi were selected using identified ROIs at the crossroads of the target bundle.
·Figure 1. Processing pipeline, from acquisitions to the atlas construction
Results:
The global tractography optimization took 18.3 hours to produce 6.7 million fibers by connecting 1.6x108 spin glasses through 2.2x1010 iterations. Tools like MRTrix or MITK face memory constraints: [8] generated 80.000 fibers from 2.5 million connected spin-glasses after 5x108 iterations within 12 to 24 hours on a single workstation. This fell short when representing the structural connectivity of the human brain. With ExaTract, we optimized 100 more spin-glasses through 100 more iterations in a comparable time frame, resulting in 84 times more fibers. We successfully recovered deep WM clusters, including the cingulum, fornix, CST, frontal aslant, uncinate bundles, corpus callosum, optical radiations, pontocerebellar fibers, anterior and posterior commissures (Fig2).
·Figure 2. Single-subject high-resolution post-mortem deep white matter atlas of the macaque using HPC global tractography
Conclusions:
ExaTract, designed for HPC and compatible with upcoming exascale supercomputers, introduces a global tractography framework free from computational constraints. It enables fast and robust fiber tracking from ultra-high resolution datasets like ex vivo mesoscale diffusion MRI but also microscale PLI datasets in the coming future. It allowed the development of a novel mesoscopic WM atlas of the Macaca fascicularis.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural)
Diffusion MRI Modeling and Analysis 1
Image Registration and Computational Anatomy
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
White Matter Anatomy, Fiber Pathways and Connectivity 2
Neuroinformatics and Data Sharing:
Brain Atlases
Keywords:
ANIMAL STUDIES
Atlasing
Computing
Data analysis
MRI
Tractography
Other - Macaque;Global tractography
1|2Indicates the priority used for review
Provide references using author date format
[2] Legeay, S. (2023) 'High-Performance Computing global tractography for ultra-high resolution diffusion MRI and 3D-PLI’, proceedings of OHBM 2023
[3] Chauvel, M. (2022), 'Investigation of the inferior fronto-occipital fasciculus in the macaque fascicularis brain using ultra-high field MRI at 11.7 T', proceedings of OHBM 2022
[4] Avants, B. B. (2009), 'Advanced normalization tools (ANTS)', Insight journal, 2(365), 1-35
[5] Karypis, G. (1997), 'METIS: A Software Package for Partitioning Unstructured Graphs, Partitioning Meshes, and Computing Fill-Reducing Orderings of Sparse Matrices', retrieved from the University of Minnesota Digital Conservancy
[6] Clarke, L. (1994), 'The MPI Message Passing Interface Standard', in Decker KM, Rehmann RM, eds. Programming Environments for Massively Parallel Distributed Systems. Monte Verità. Birkhäuser, pp. 213-218
[7] Fillard, P. (2009), 'A Novel Global Tractography Algorithm Based on an Adaptive Spin Glass Model', Medical Image Computing and Computer-Assisted Intervention − MICCAI 2009, pp. 927-934
[8] Reisert, M. (2009), 'Global reconstruction of neuronal fibres', In proceeding of MICCAI Diffusion Modelling Workshop 2009
[9] Chauvel, M. (2023), 'In vivo mapping of the deep and superficial white matter connectivity in the chimpanzee brain', NeuroImage, vol. 282
[10] https://framagit.org/cpoupon/gkg