FreeSurfer Development, Maintenance, and Hardening

Massachusetts General Hospital, Boston MA

Investigators

Linked publications, trials & patents

Abstract

Imaging of the human brain has seen explosive growth in the last two decades mainly through the various modalities of MRI. The massive amount of data requires automatic and robust tools for analysis. FreeSurfer (FS, surfer.nmr.mgh.harvard.edu) is one of the preeminent tools used for neuroimage analysis. FS has more than 76,000 downloads, and the core FS manuscripts have been cited more than 50,000 times. FS is part of the analysis core for many NIH-funded large-scale data acquisition projects such as the Human Connectome Project (HCP), Alzheimer's Disease Neuroimaging Initiative (ADNI), Framingham Heart Study (FHS), The Adolescent Brain Cognitive Development (ABCD), as well as the UK BioBank. One third of the 600+ ADNI-based publications cite FS. Simply put, much of the innovative research done in neuroimaging would not be possible without FS. Started in 1998, FS is best known for providing detailed and automated anatomical analysis of T1-weighted MRI images, especially for the cortical surface. However, FS anatomical analysis provides an ideal substrate for all modes of brain imaging including functional MRI, diffusion MRI, PET, optical/NIRS, as well as EEG/MEG. FS provides tools to perform these analyses as well as software to integrate with other analysis tools (e.g., SPM, FSL, AFNI). FS has been used for presurgical planning and even in the operating room. In this renewal, we are proposing several significant and innovative projects. The first project is to create a High- Resolution Surface-based Cytoarchitectonic Cortical Label Atlas. For this we have teamed up with Dr. Katrin Amunt’s group at Juelich University. Over the past 20 years, Dr. Amunts’ group has created 130 cytoarchitectonically defined cortical labels (per hemisphere) on 23 ex vivo cases. The labels are defined in the volume. Here we aim to create cortical surface mesh models of the 23 ex vivo, sample the cytoarchitonic labels onto the surface, then create a surface-based probabilistic atlas which can then be applied to in vivo cases to render a segmentation of cortex in stunning detail. The second is Joint surface-based registration and atlas construction of brain geometry and function (JOSA). JOSA uses deep learning to perform surface-based registration between subjects. During training, JOSA takes cortical geometry (eg, curvature) as input as well as an auxiliary feature set (AFS). The network does not have direct access to the AFS, but the AFS influences the network through a semi-supervised loss term during training, which forces the network to learn to predict and align the AFS from only geometric information. The result can be used to aligned AFS-defined regions (eg, fMRI-defined language areas) without using the AFS at inference, making it suitable in situations when features are difficult or impossible to obtain at application time, such as parcellations, architectonic ROIs, transcriptomics, or molecular profiles. JOSA can be used to leverage the thousands of fMRI, DTI, PET, and ex vivo data sets to create new atlases. The third project is to improve within-subject cortical surface placement for longitudinal studies to reduce within-subject noise. In addition to this new technical development, we are requesting support for software engineering, maintenance, and user support – mundane and not innovative, but high-impact this type of support is critical to the thousands of researchers who rely on FreeSurfer.