Continued Development of Infant Neuroimaging Analysis Tools

The first two years are an exceptionally dynamic and critical period of brain development, featuring significant growth in both cerebrum and cerebellum. The availability of large-scale, multi-site infant MRI datasets, e.g., the developing Human Connectome Project (dHCP), Baby Connectome Project (BCP), and National Database for Autism Research (NDAR), affords unprecedented opportunities for precise charting the dynamic early brain development, providing important insights into the origins and aberrant growth trajectories of neurodevelopmental disorders, such as autism. However, existing neuroimage analysis tools designed for adults are not suitable for infant neuroimages. In 2020, our team has successfully developed and released iBEAT V2.0 (infant Brain Extraction and Analysis Toolbox) with advanced deep learning techniques, which has successfully processed 31,000+ infant scans from 200+ institutions with diverse imaging protocols and scanners. iBEAT has directly contributed to 50+ journal publications, including Brain, Nature Methods, Neuron, Nature Communications, PNAS, Neuroimage, and Cell Reports. However, iBEAT still has two major limitations. 1) It focused on the infant cerebrum MRIs and thus is inapplicable for the more challenging cerebellum MRIs, which exhibit much thinner and more tightly folded cortex than the cerebrum, extremely low and dynamic tissue contrast, and suffer from large domain-shift issue across imaging sites. 2) Certain important functionalities for cerebrums are either missing, e.g., motion correction, subcortical segmentation, volumetric parcellation, and surface registration, or have degraded performance in certain scenarios. To address these issues, this project aims to significantly enrich iBEAT by 1) creating deep learning-based computational tools for cerebellar tissue segmentation, atlas building, surface reconstruction and parcellation, as all as 2) adding new cerebrum-related functionalities and significantly improving existing functionalities, to enable comprehensive, accurate, and integrative analysis of cerebrum and cerebellum and their interplay during infancy. Accordingly, we propose five specific aims. Specifically, we will develop a novel prior-guided cerebellum tissue segmentation method with self-verification (Aim 1). We will then construct the first 4D infant cerebellum atlases with longitudinally consistent, temporally continuous, and spatially detailed patterns, by developing a novel unsupervised learning-based anatomy-guided atlas construction framework (Aim 2). We will reconstruct topologically correct and geometrically accurate cerebellar cortical surfaces and further develop a novel Spherical Surface Transformer to precisely parcellate cerebellar cortical surfaces into anatomically meaningful regions (Aim 3). We will add new cerebrum-related modules for motion correction, subcortical segmentation, volumetric parcellation, and surface registration, and further improve existing modules in terms of robustness and accuracy with our new techniques (Aim 4). Finally, we will undertake a comprehensive upgrade for iBEAT, by improving usability, robustness, compatibility, code structure, and documentation (Aim 5).