Title: Computational Neuroimaging MRI for Studying Early Brain Development with Autism
Due to the absence of early biomarkers of autism, diagnosis must rely on behavioral observations long after birth,
leading to missed opportunities for early intervention. Thus, it is of great importance to detect autism earlier in life
for better intervention. The increasing availability of large-scale multimodal infant neuroimaging data
(structural, functional and diffusion MRIs), e.g., Multi-visit Advanced Pediatric (MAP) Brain Imaging Study, Baby
Connectome Project (BCP), and National Database for Autism Research (NDAR), affords unprecedented
opportunities for precise charting of dynamic early brain developmental trajectories of autism, potentially providing
important clues relevant to early detection of autism. Our hypothesis is that accurate characterization
(segmentation, parcellation, multimodal neuroimaging measurements) of infant brain MRIs acquired from multiple
centers will provide important insights into the origins and aberrant growth trajectories of autism and help identify
potential imaging-based biomarkers for the early diagnosis of autism.
To fully benefit from the large-scale multi-center infant neuroimaging data, a major barrier is the critical lack
of computational characterization tools for accurate and robust processing of the challenging infant MRIs,
which typically exhibit dynamic and extremely low tissue contrast, and large data heterogeneity. Building upon
our existing research, this project aims to create and disseminate novel computational characterization tools
that will enable accurate segmentation of cerebrum, cerebellum, and subcortical structures, parcellation, and
measurements of infant brain structural, functional and diffusion MRIs from multiple imaging centers, and to
further identify imaging-based biomarkers for early diagnosis of at-risk infants. To achieve our goal, we propose
4 specific aims. We will develop a novel contrast-enhancement network to increase the tissue contrast and a
novel prior-guided transformer for enhanced and harmonized cortical and subcortical segmentation (Aim 1). We
will propose a novel joint super resolution and tissue segmentation for cerebellum, aiming to achieve fine-grained
segmentation results in an isotropic 0.4mm (or higher) space (Aim 2). Based on essential semantic features from
segmentation maps in Aims 1 and 2, as well as contextual folding features from cortical surfaces, we will further
develop a novel hybrid volume-surface parcellation framework trained on an innovatively augmented large-scale
and diverse dataset (Aim 3). Finally, we develop a joint clinical scores regression and diagnosis model with
attention mechanisms using multimodal features, including volumetric features from T1w and T2w scans,
segmentation and parcellation maps (Aims 1-3), surface features from subcortical structures, cerebral and
cerebellar cortex, and connectivity features from fMRI and dMRI, aiming to achieve high accuracy and
interpretability in autism diagnosis (Aim 4). We will freely release our tools and all our processed data, to NDAR.