Conditional Fetal Brain Atlas Learning for Automatic Tissue Segmentation

TL;DR

This work tackles the variability in fetal brain development and imaging by proposing a continuous, age-conditioned fetal brain atlas that enables real-time tissue segmentation. It combines a FiLM-conditioned Template Generation Network, a U-Net-based Registration Network, and a projection discriminator in an end-to-end adversarial framework to produce age-specific templates and subject-specific warps. The approach is evaluated on 219 neurotypical fetal MRIs (21–37 weeks) and achieves competitive atlas quality and segmentation accuracy (DSC around 86.3% for six tissues) with demonstrable GA-dependent anatomical fidelity and real-time performance. Overall, the method offers a practical tool for personalized fetal brain assessment with potential clinical and research impact, including detailed neurodevelopmental trajectories.

Abstract

Magnetic Resonance Imaging (MRI) of the fetal brain has become a key tool for studying brain development in vivo. Yet, its assessment remains challenging due to variability in brain maturation, imaging protocols, and uncertain estimates of Gestational Age (GA). To overcome these, brain atlases provide a standardized reference framework that facilitates objective evaluation and comparison across subjects by aligning the atlas and subjects in a common coordinate system. In this work, we introduce a novel deep-learning framework for generating continuous, age-specific fetal brain atlases for real-time fetal brain tissue segmentation. The framework combines a direct registration model with a conditional discriminator. Trained on a curated dataset of 219 neurotypical fetal MRIs spanning from 21 to 37 weeks of gestation. The method achieves high registration accuracy, captures dynamic anatomical changes with sharp structural detail, and robust segmentation performance with an average Dice Similarity Coefficient (DSC) of 86.3% across six brain tissues. Furthermore, volumetric analysis of the generated atlases reveals detailed neurotypical growth trajectories, providing valuable insights into the maturation of the fetal brain. This approach enables individualized developmental assessment with minimal pre-processing and real-time performance, supporting both research and clinical applications. The model code is available at https://github.com/cirmuw/fetal-brain-atlas

Figures (5)

• Figure 1: The model architecture of the proposed conditional atlas learning network incorporates two networks: 1.) The template generation framework, including age-appropriate structural representation and its associated anatomical labels. 2.) The U-net-based registration framework with the additional projection discriminator.
• Figure 2: a) The histogram illustrates the heterogeneous age distribution of the neurotypical dataset divided in 1.5 and 3 Tesla acquisitions. b) Exemplary SVR samples across different GA.
• Figure 3: a) Axial slices of the generated templates from the baseline approaches and our best-performing model. b) Visualization of the generated template (Atlas), test subject (Fix), warped atlas (Warp), and associated segmentation maps for a brain at 35 GW. The red box highlights a region with pronounced individual cortical folding.
• Figure 4: a) Neurodevelopmental trajectory of ventricles volume for training dataset (solid line), generated atlas (squares), test labels (triangle), and predictions (circle). b) Radar plot illustrating segmentation performance across different GWs. Each axis represents an anatomical label and the distance indicates the DSC.
• Figure S1: Volumetric trajectories of eCSF, cGM, total WM, brainstem, and deep GM, (from top left to bottom right) from 21 to 37 GW. The trajectory of the training data, including its standard deviation, is modeled using a polynomial fit. The brain region-specific volumes are shown for the test labels ($\bigtriangleup$) and the predicted values ($\bigcirc$).