Cross-scale integration and modeling

Massachusetts General Hospital, Boston MA

Investigators

Linked publications & trials

Abstract

Cross-scale integration and modeling No current imaging technology can directly and without significant distortion visualize the defining three dimensional microscopic features of the human brain. Ex vivo histological techniques yield exquisite planar images, but the cutting, mounting and staining they require induce slice-specific distortions, introducing cross- slice differences that prohibit true 3D analysis. Clearing techniques such as CLARITY have proven difficult to apply to large blocks of human tissue and cause dramatic distortions as well. Thus, we have only a poor understanding of human brain structures that occur at a scale of 1-200μm, in which neurons are organized into functional cohorts. This impairs our ability to understand brain function, as the functional properties of any given cell are a function of both its molecular characteristics and the spatial context within which it resides. To date, mesoscopic features such as cortical laminar and columnar structures that are critical components of this spatial context, have only been quantified in studies of 2D histologic images acquired in a small number of subjects and/or over a small region of the brain, typically in the coronal orientation, implying that features that are oblique or orthogonal to the coronal plane are difficult to properly analyze. In TRD1 we aim to develop analysis tools to overcome many of these barriers and make micro- and mesoscopic information available in commonly used coordinate frameworks. In Aim 1 we seek to improve preprocessing pipelines for ex vivo Optical Coherence Tomography (OCT) and MRI, including measuring and removing mesoscale distortions and leveraging deep learning (DL) to remove noise and artifacts to improve the quality of the images, in collaboration with CP Akkin (OCT) and SP Tisdall (MRI). In Aim 2 we will develop the infrastructure to map micro- and mesoscale images to whole-brain coordinate systems, enabling the transfer of information from e.g. light-sheet microscopy to MRI. This will be accomplished through the development of a set of innovative algorithms – one that segments vascular trees in ex vivo images and one that incorporates segmentations into a generative model - allowing the automated creation of large training sets in novel imaging modalities without requiring manual labeling. This novel functionality will be of particular importance as it will open up the vessel segmentation to any imaging modality in which vessels are visible without the need for manual annotations in the target modality. The vascular segmentation will then serve as endogenous fiducial markers to guide cross-modal, cross-scale registration algorithms in collaboration with CP Gee. Finally, in Aim 3 we will use the higher-quality images from Aim 1 and the registration from Aim 2 to build models of mesoscopic structures such as cortical layers and develop cross-subject registration tools that optimally warp them into common coordinate systems, making them easily available to the larger neuroimaging community.