Research

In the biomedical research community, there exists a need for novel tools and strategies to more precisely model the brain’s structure and function. New techniques can enhance the current understanding of the developing brain and healthy adult brain, as well as various disease states, transcending traditional boundaries such as age, sex and species.

LONIR furthers this mission by developing resources to aid investigators and by facilitating collaboration between researchers around the world. By creating and sharing a diverse array of tools to analyze and visualize imaging data, LONIR fosters the interdisciplinary partnerships that are vital to the advancement of brain research.

Now in its nineteenth year, LONIR aims to continue providing innovative solutions for the investigation of imaging, genetics, behavioral and clinical data. Projects have been designed within the fields of Technology, Research, and Development (TR&D).

Research Areas

Data Science

Data Science focuses on the methodological developments for the analysis of brain imagery. Specifically, this project will design and distribute new methods for robust image segmentation and registration, quality assurance and evaluation of image processing results, and processing of structural and diffusion brain data.

Diffusion MRI and Connectomics

Diffusion MRI and Connectomics will advance the study of brain connectivity using diffusion imaging and its powerful extensions. This project will use Deep Learning to develop tract-based statistical analysis tools and adaptive connectivity mapping approaches. It will enable analysis of large diffusion imaging datasets totaling over 10,000 subjects.

Intrinsic Surface Mapping

Intrinsic Surface Mapping develops novel algorithms for surface reconstruction, modeling and analysis. By removing unnecessary metric distortion, it will provide greatly improved accuracy and power in detecting alterations to brain anatomy and function in disease studies.

Current Research

The Data Science core, or TR&D1, is working to quantify, control, and monitor image quality as an essential prerequisite for ensuring the validity and reproducibility of many types of neuroimaging data analyses. However, when performing large cohort neuroimaging studies, manual work for these tasks becomes nearly impossible.

TR&D1 has developed machine learning and deep learning approaches for automation of MR image QC (quality control), such as LONI-QC and a motion artifact correction tool. This work is now being applied to QC and image enhancement for other image-modality data, including diffusion, functional and perfusion imaging.

 
The Diffusion MRI and Connectomics core, or TR&D2, has focused on improving models of tissue microstructure using tensor distribution function (TDF) and multi-shell methods, automated extraction and statistical analysis of white matter tracts and fiber bundles, and new brain connectivity analysis methods. Here are a few highlights:

Generative Adversarial Networks for Diffusion MRI Harmonization: TR&D2 created a method to harmonize diffusion MRI data from multiple scanners and sites that uses generative adversarial networks to produce “scanner invariant representations.” A variational autoencoder learns an intermediate representation of the data that is invariant to the scanning protocol, in a technique adapted from information theory-based algorithmic fairness.

The intermediate representation is then used to create an image reconstruction that is uninformative of its original source, but still faithful to underlying structures. This method was on paired data from the 2018 MICCAI Computational Diffusion MRI (CDMRI) Challenge Harmonization dataset and it outperformed a recent method in mapping data from three different scanning contexts to and from one separate target scanning context.

Quantitative Imaging Toolkit (QIT) for Diffusion MRI Tractography: TR&D2 also developed, disseminated, and offered training in a comprehensive software toolkit for diffusion MRI tractography. QIT is a software package of computational tools to model, analyze, and visualize scientific imaging data. It was developed for tractography and microstructure analysis of diffusion MRI datasets, but also supports many different data types, including multi-channel volumetric data, multi-label masks, curves, triangle meshes, geometric primitives, tabular data, and spatial transformations.

QIT provides an application called qitview for interactive 3D rendering and data analysis, as well as command-line tools to do batch processing and scripting. QIT also offers ways to integrate these tools into grid computing environments and scientific workflows, such as the LONI Pipeline.

 

The Intrinsic Surface Mapping core, or TR&D3, has extended its surface mapping method based on Riemannian Metric Optimization on Surfaces (RMOS) in the Laplace-Beltrami (LB) embedding space, developed a series of tools for surface-based tractography and 3D surface reconstruction and mapping, and developed novel mathematical algorithms for the analysis of topographic regularity in fiber tracts from diffusion MRI. Here are a few highlights:

Patch-based RMOS for Alzheimer’s Disease (AD) Imaging Research: TR&D3 extended its intrinsic surface mapping method based on RMOS in the LB embedding space to the mapping of surface patches on cortices, demonstrating that the transentorhinal region can be more accurately segmented than using the spherical registration in FreeSurfer. In this project period, TR&D3 continued the development and validation of its patch-based RMOS method on large-scale and multi-modal imaging data from AD imaging research.
Retinal Microvasculature Reconstruction and Analysis: Using the LB spectrum, TR&D3 has developed a series of tools for 3D surface reconstruction and mapping. While these tools were developed in the context of brain imaging, many of them are generally applicable to other medical imaging data. TR&D3 successfully applied them to model and analyze the retinal microvasculature based on optical coherence tomography angiography (OCTA).

 
The Administration core upgraded its storage resources in order to support the growth of virtual infrastructure for Docker containers, virtual machines, etc. The Dell ECM Unity XT 480F All-Flash Storage unit has already become an essential resource that helps all of the Institute’s compute and storage operations work faster and more reliably. Additionally, the powerful new Dell EMC Isilon A2000 NAS Storage cluster, with 52 nodes and an additional 10Pb of usable storage capacity, brings the total storage capacity of the Institute above 18Pb of highly available, high-performance storage.

The Training and Dissemination core hosted several new workshops at the Institute, including the first annual SoCal High-Lo Field MR Imaging Workshop, which convened researchers from the imaging community across Southern California. It also launched new partnerships with visiting scholars, undergraduate degree programs, and high school training initiatives. Visit the Training & Dissemination page to learn more.

Methods

The approach to integrated computational neuroscience is based on a scalable, portable and distributed infrastructure, which uses object-oriented programming, Extensible Markup Language (XML), encrypted distributed computing, and open-source design, implementation, and tool dissemination.

Algorithms are designed to generate average models of brain anatomy and maps of growth, degeneration and their population statistics. These models are based on parametric surfaces, volumetric morphology, and topology-preservation mapping.

The resulting algorithms are implemented, validated, and distributed via the LONIR Pipeline environment and are applicable to a variety of computational neuroscience challenges in normal brain and disease.

Other Projects

The LONIR structure is designed to facilitate studies of dynamically changing anatomic frameworks: developmental, neurodegenerative, traumatic and metastatic. LONIR therefore targets new strategies for surface and volume parameterization that track change over time. Additional research cores include anatomic fundamentals and analyzing anatomic and cytoarchitectural attributes across multiple spaces and time.

LONIR also has a core that focuses on visualization and animation, which creates and distributes brain models that depict complex variations in brain structure and function over time. This includes developing projects for handling cortical data.

Ongoing national and international collaborations also encompass a diverse array of research foci, including Alzheimer’s disease, traumatic brain injury, epilepsy, autism, HIV, blindness, brain development, and connectivity.