The 4D Human Embryonic Brain Atlas: spatiotemporal atlas generation for rapid anatomical changes

TL;DR

This work introduces the 4D Human Embryonic Brain Atlas, a first spatiotemporal ultrasound-based atlas covering days 56–90 of gestation to model rapid first-trimester brain development. A two-network deep learning framework generates an atlas A_t from a time-dependent initial atlas A_t^0 and learns nonrigid deformations to map individual images I_{i,t} to the atlas, using a loss that includes a novel atlas-deviation term to preserve age-specific anatomy. The atlas is built from 831 ultrasound images in the Rotterdam Periconceptional Cohort, validated via ablation studies, visual comparisons to ex vivo and fetal atlases, and a VBM analysis showing sensitivity to maternal BMI effects on brain morphology. The 4D atlas enables quantitative analysis of early brain development, supports early anomaly detection, and provides an engaging visualization tool for clinicians and expectant parents, potentially informing interventions and research in prenatal neurodevelopmental health.

Abstract

Early brain development is crucial for lifelong neurodevelopmental health. However, current clinical practice offers limited knowledge of normal embryonic brain anatomy on ultrasound, despite the brain undergoing rapid changes within the time-span of days. To provide detailed insights into normal brain development and identify deviations, we created the 4D Human Embryonic Brain Atlas using a deep learning-based approach for groupwise registration and spatiotemporal atlas generation. Our method introduced a time-dependent initial atlas and penalized deviations from it, ensuring age-specific anatomy was maintained throughout rapid development. The atlas was generated and validated using 831 3D ultrasound images from 402 subjects in the Rotterdam Periconceptional Cohort, acquired between gestational weeks 8 and 12. We evaluated the effectiveness of our approach with an ablation study, which demonstrated that incorporating a time-dependent initial atlas and penalization produced anatomically accurate results. In contrast, omitting these adaptations led to anatomically incorrect atlas. Visual comparisons with an existing ex-vivo embryo atlas further confirmed the anatomical accuracy of our atlas. In conclusion, the proposed method successfully captures the rapid anotomical development of the embryonic brain. The resulting 4D Human Embryonic Brain Atlas provides a unique insights into this crucial early life period and holds the potential for improving the detection, prevention, and treatment of prenatal neurodevelopmental disorders.

Figures (16)

• Figure 1: Overview of the proposed method and network architectures for spatiotemporal atlas generation. The atlas generator network takes as input the gestational age $t$ (in days, normalized between 0 and 1) and the initial atlas $A_t^0$, and outputs the generated atlas $A_t = A_t^0 + A_t^g$. It consists of a dense layer followed by three upsampling convolutional layers with 32 filters, with the number of upsampling layers chosen such that the final resolution matches that of the input images $I_{i,t}$. A final convolutional layer at full resolution produces $A_t^g$. The nonrigid registration network follows the VoxelMorph architecture Balakrishnan2019 and includes an encoder with one convolutional layer of 16 filters and three layers with 32 filters (stride 2), a decoder with five upsampling convolutional layers with 32 filters, skip connections, and two convolutional layers with 16 filters at full resolution. Diffeomorphic deformations are obtained using a stationary velocity field representation, where the deformation field $\phi_{i,t}$ is computed by integrating the velocity field $\nu_{i,t}$, and the inverse deformation $\phi_{i,t}^{-1}$ by integrating its negation $-\nu_{i,t}$Ashburner2007.
• Figure 2: Dataset characteristics. (a) The number of ultrasound scans available per gestational day for the atlas-generation set, validation set, test set and in total. (b) The number of subjects with $1,2, \ldots, 5$ ultrasound acquisitions. (c) The number of subjects per BMI category in the atlas-generation, validation and test set, where the categories are defined as follows: low: BMI < 19, normal: 19 $\leq$ BMI < 25, overweight: 25 $\leq$ BMI < 30, and obese: BMI $\geq$ 30.
• Figure 3: Example of mid-sagittal, mid-axial and mid-coronal plane at different gestational days after preprocessing of the imaging data. Brain structures are marked with arrows. LV = lateral ventricle, 3V = third ventricle, M = cavity of the mesencephalon, 4V = fourth ventricle, CP = choroid plexus, CER = cerebellum.
• Figure 4: Atlas $A_t$ at different gestational days for different hyperparameter values tested in Experiment 1. All images are shown in the mid-sagittal plane. The arrows indicate relevant visible brain structures. LV = lateral ventricle, 3V = third ventricle (cavity of the diencephalon), M = cavity of the mesencephalon, 4V = fourth ventricle (cavity of the rhombencephalon), CP = choroid plexus, CER = cerebellum.
• Figure 5: HV-curve for $\delta=3$ in Experiment 1. The gray line is the reference HV-curve $\text{HV}_{\text{VR}}(t)$ derived in previous research Koning2016 and the dots represent $V(A_t)$ the HV for every gestational day of the atlas $A_t$.