Recent advances in functional magnetic resonance imaging (fMRI) have helped elucidate previously inaccessible trajectories of early-life prenatal and neonatal brain development. To date, the interpretation of fetal-neonatal fMRI data has relied on linear analytic models, akin to adult neuroimaging data. However, unlike the adult brain, the fetal and newborn brain develops extraordinarily rapidly, far outpacing any other brain development period across the lifespan. Consequently, conventional linear computational models may not adequately capture these accelerated and complex neurodevelopmental trajectories during this critical period of brain development along the prenatal-neonatal continuum. To obtain a nuanced understanding of fetal-neonatal brain development, including non-linear growth, for the first time, we developed quantitative, systems-wide representations of brain activity in a large sample (>500) of fetuses, preterm, and full-term neonates using an unsupervised deep generative model called Variational Autoencoder (VAE), a model previously shown to be superior to linear models in representing complex resting state data in healthy adults. Here, we demonstrated that non-linear brain features, i.e., latent variables, derived with the VAE pretrained on rsfMRI of human adults, carried important individual neural signatures, leading to improved representation of prenatal-neonatal brain maturational patterns and more accurate and stable age prediction in the neonate cohort compared to linear models. Using the VAE decoder, we also revealed distinct functional brain networks spanning the sensory and default mode networks. Using the VAE, we are able to reliably capture and quantify complex, non-linear fetal-neonatal functional neural connectivity. This will lay the critical foundation for detailed mapping of healthy and aberrant functional brain signatures that have their origins in fetal life.

This dataset includes 270 processed resting state fMRI scans from 95 healthy fetuses and 160 full-term born healthy infants. Scans were collected as a part of an ongoing longitudinal study examining prenatal-neonatal brain development at Children’s National Hospital in Washington DC. Fetuses (49 female) from healthy pregnancies were scanned between 19.14–39.71 gestational weeks (mean and SD; 33.65±4.01). Postmenstrual age of full-term born infants at scan ranges between 37.71–47.43 weeks (41.67±1.84). Maternal exclusion criteria were psychiatric disorders, metabolic disorders, genetic disorders, complicated pregnancies, multiple pregnancies, alcohol, and tobacco use, maternal medications, and contraindications to MRI. All experiments were conducted under the regulations and guidelines approved by the Institutional Review Board (IRB) of Children’s National; written informed consent was obtained from each pregnant woman who participated in the study.

For fetal scans, structural and functional resting-state MR images were acquired using a 1.5 Tesla GE MRI scanner with 8-channel receiver coil. The structural MR images for the fetal brain were acquired using single-shot fast spin-echo T2-weighted images by following settings: TR=1100ms, TE=160ms, flip angle=90 degrees, and voxel size= 0.8 0.8 2mm. Functional data were acquired using echo planar images (EPI) with TR=3000ms, TE=60ms, flip angle=90 degrees, field of view= 33cm, and voxel size = 2.58 2.58 3mm, and total scan volume=144 (=7.2min). The structural and functional MR images of full-term infant brain MRI studies were acquired using 3T GE scanner. T2-weighted fast spin echo MRI was obtained using the following parameters: TR=2500ms; TE=64.49ms, voxel size= 0.625 1 0.625mm. The parameters of fMRI MRI scans were set to TR=2000ms, TE=35ms, voxel size=3.125 3.125 3mm, flip angle= 60 degrees, field of view=100mm, and total scan volume=200~300 (=6.7~10min).

Functional MR images were preprocessed as followings: slice time correction, bias-field correction, motion-correction, spatial smoothing at full-width-half-maximum=4.5mm, and intensity scaling. We excluded volumes having excessive head motion (frame-wise motion >1mm or rotational motion >1.5 degrees). Neonatal MRI scans with < 4 mins were excluded in the analysis. The analyzed data length of fetus and neonate group was 4-7 (mean S.D.=5.4 0.9) and 4-8.9 (5.5 0.8) mins, respectively. Finally, preprocessed rsfMRI scans at the volumetric brain space were projected to the standard cortical space using HCP workbench command -volume-to-surface-mapping.

This dataset contains three types of latent variables from preprocessed fMRI data using different types of models (VAE, ICA, and cortical parcel). The details of VAE can be found at Kim, Jung-Hoon, et al. "Representation learning of resting state fMRI with variational autoencoder." NeuroImage 241 (2021): 118423. Actual implementation of code can be found at GitHub - libilab/rsfMRI-VAE: Pytorch implementation of 'Representation Learning of Resting State fMRI with Variational Autoencoder'. Details regarding ICA map and cortical parcel can be found at www.humanconnectome.org.

The format of the dataset is TXT file. Columns and rows of each text file stand for latent variables (for example, 512 for VAE) and time points (variable over different subjects). Any type of program will be compatible with this dataset but we recommend using MATLAB. 

For the latent variable from VAE, there are 512 latent variables: first 256 and later 256 stand for the mean and log variance of 256 latent variables, respectively. For ICA, there are different number of latent variables (50, 100, 200, and 300) based on which ICA maps were used. For cortical parcel, there are 360 latent variables. 

If you have any questions, please contact the first author (Junghoon Kim, PhD) or the corresponding author.