BOBs (Baby Open Brains) Repository: An Open-Science Repository of Segmentations for Human Infants

• In Event: Functional and structural findings from the Baby Connectome Project

Abstract

Introduction: Longitudinal early-life neuroimaging studies, like the Healthy Brain Cognitive Development (HBCD) study (Volkow, 2021), require processing pipelines that can work across the human lifespan. In particular, efforts to perform surface-based analysis require the extension of processing tools like Freesurfer (Fischl, 2012) into the infant domain (Zollei, 2020). Efforts toward such extensions remain limited by the availability of manually-corrected and Freesurfer-compliant segmentation anatomical MRI data (Rodrigues, 2015). High quality segmentations require considerable effort and expertise, and would therefore benefit from review in an open repository. Therefore, we constructed the Baby Open Brains (BOBs) repository. BOBs repository comprises a set of Freesurfer-compatible human infant brain segmentations that are manually curated and expert reviewed. The entire community can view, comment, edit, and improve. Datalad (Halchenko et al., 2021) is used to version control BOBs repository and ensure data provenance. OpenNeuro (Markiewicz et al., 2021) hosts BOBs repository for accessibility. BrainBox (Heuer et al., 2016) can be used to both view, comment, and refine segmentations, enabling continuous improvement.
Methods: The Baby Connectome Project (BCP) was curated based on anatomic data quality (n=80 total) of which 73 infants aged 0-8 months were selected for data processing. All data were pre-processed using the DCAN-infant pipeline, and the spatially normed MNI Infant, AC-PC-aligned T1 and T2 were used to guide segmentations. Initial segmentations were constructed in one of two ways, either using ANTS joint label fusion (JLF; Wang, 2012), or using the BIBSnet algorithm (Hendrickson, 2022) trained on a subset of the final segmented data. Initial segmentations were then provided to raters, who were guided by an expert rater to correct the white matter and gray matter for the cerebral hemispheres for each segmentation. All segmentations were reviewed and approved by the expert rater prior to upload. Segmentations were version controlled via datalad, and then uploaded to OpenNeuro, and can be viewed, commented, and edited using BrainBox. A subset of JLF and manual segmentations (n = 38) were then re-run through the DCAN-infant pipeline and the output volumes were evaluated .
Results: Manual corrections show substantial improvement over prior available approaches for automated infant segmentations. In particular, comparisons of T1/T2 identified unmyelinated white matter and gray matter shows substantially better annotation across infant age ranges. JLF volumes ,compared to manually corrected volumes, show significantly inflated gray matter (T=4.68, p<0.001, DSC=0.868) and deflated white matter (T=8.13, p<0.001, DSC=0.85) estimates (see: Figure 1).
Conclusions and Future Directions: The BOBs repository should help improve researchers’ ability to construct, improve, and test infant pipelines that incorporate freesurfer-compliant segmentations. In addition, we will continue to refine our 0-8 month segmentations, while extending into subsequent 1, 2, and 3 year olds. Studies like HBCD will find such repositories critical for ensuring best standards and practices for data processing.

Authors

• Eric Feczko, University of Minnesota
  Presenting Author

• Sally Stoyell, University of Minnesota - Twin Cities

• Lucille A Moore, University of Minnesota - Twin Cities

• Timothy Hendrickson, University of Minnesota

• Paul Reiners, University of Minnesota

• Anders Perrone

• Dimitrios Alexopoulos

• Omid Kardan

• Taylor Chamberlain

• Anurima Mummaneni, University of Chicago

• Henry Caldas

• Tabitha Martin, University of Minnesota

• Sooyeon Sung, University of Minnesota - Twin Cities

• Alexandra Harper, University of Minnesota

• Ekomobong Edet Eyoh, University of Minnesota - Twin Cities

• Sarah Grace Kuplic

• Katelyn Day, University of Minnesota

• Trevor K.M. Day, University of Minnesota - Twin Cities

• Nora Byington

• Gracie Grimsrud, University of Minnesota

• Kenneth Barrett, PrimeNeuro

• BettyAnn Chodkowski, Vir Biotechnology, Inc.

• Brayton Hall, PrimeNeuro

• Dhruman Goradia, PrimeNeuro

• Amanda R. Rueter, University of Minnesota

• Christopher D. Smyser, Washington University School of Medicine

• Monica Rosenberg, University of Chicago

• Alice M. Graham, Oregon Health and Science University

• Martin Styner, University of North Carolina at Chapel Hill

• Damien A. Fair, University of Minnesota - Twin Cities

• Jed T. Elison, University of Minnesota - Twin Cities