Automated Segmentation Techniques in Neonatal Brain Imaging

Automated segmentation of neonatal brain MRI has become a cornerstone in the study of early neurodevelopment, enabling large-scale quantification of tissue growth, cortical folding and regional maturation. Neonatal brains present unique challenges: rapidly evolving tissue contrast, small structure sizes, motion artefacts and wide inter-site variability in scanner protocols. Traditional atlas-based approaches rely on labelled templates propagated to individual scans, but these can struggle with the high heterogeneity of neonatal data. More recent methods have harnessed machine learning and deep neural networks to capture complex intensity patterns and shape priors, while self-supervised frameworks exploit unlabelled data to refine segmentation boundaries without extensive manual annotation. Multi-atlas label fusion techniques improve robustness by combining segmentations from multiple templates. Advances in cortical surface reconstruction, cerebellar parcellation and tissue-type classification now allow high-resolution mapping of grey and white matter development, with outcomes that correlate with neurodevelopmental trajectories in preterm and term‐born infants. Robust automated pipelines are increasingly deployed to detect subtle deviations associated with prematurity, congenital anomalies and early neuropsychiatric risk, supporting early intervention strategies and the creation of normative growth atlases across populations and MRI platforms.

Research from Nature Portfolio

A novel self-supervised learning framework for infant cerebellum segmentation has demonstrated high accuracy across three large datasets, capturing the rapid growth patterns of grey and white matter in the first six months of life. This approach reduces reliance on manual labels by leveraging consistency across unpaired scans, and reveals sex and condition-specific differences in cerebellar volume, paving the way for large-scale studies of typical and atypical neurodevelopment. Another key development is a sparsity-based atlas selection and machine-learning label fusion method for neonatal brain extraction. By selecting a minimal but representative set of atlases and employing a learning-based fusion strategy, this pipeline outperformed eleven publicly available extraction tools in multi-modal datasets, offering scalable segmentation that adapts to partial labelling and diverse imaging protocols.

Automated Segmentation Techniques in Neonatal Brain Imaging publication trend

The graph below shows the total number of articles in automated segmentation techniques in neonatal brain imaging across all publications each year (not limited to Nature Index journals).

Technical terms

Atlas-based segmentation: A method that propagates labelled brain templates to new scans to delineate anatomical regions.

Deep learning: A machine learning approach using layered neural networks to learn complex image features for segmentation.

Label fusion: The combination of outputs from multiple atlases or classifiers to produce a consensus segmentation.

Dice coefficient: A statistical measure of overlap between automated and reference segmentations, ranging from 0 (no overlap) to 1 (perfect overlap).

Self-supervised learning: A training strategy that uses intrinsic data properties rather than manual labels to learn useful image representations for segmentation.

References

1. Self-supervised learning with application for infant cerebellum segmentation and analysis. Nature Communications (2023).
2. Cortical growth from infancy to adolescence in preterm and term-born children. Brain (2023).
3. Neonatal Brain Tissue Classification with Morphological Adaptation and Unified Segmentation. Frontiers in Neuroinformatics (2016).
4. Accurate Learning with Few Atlases (ALFA): an algorithm for MRI neonatal brain extraction and comparison with 11 publicly available methods. Scientific Reports (2016).