Magnetic Resonance Imaging of the Newborn Brain: Automatic Segmentation of Brain Images into 50 Anatomical Regions

Ioannis S. Gousias, Alexander Hammers, Serena J. Counsell, Latha Srinivasan, Mary Rutherford, Rolf A. Heckemann, Joseph V. Hajnal, Daniel Rueckert, A. David EdwardsPLoS 2013 · DOI

Abstract

We studied methods for the automatic segmentation of neonatal and developing brain images into 50 anatomical regions, utilizing a new set of manually segmented magnetic resonance (MR) images from 5 term-born and 15 preterm infants imaged at term corrected age called ALBERTs. Two methods were compared: individual registrations with label propagation and fusion; and template based registration with propagation of a maximum probability neonatal ALBERT (MPNA). In both cases we evaluated the performance of different neonatal atlases and MPNA, and the approaches were compared with the manual segmentations by means of the Dice overlap coefficient. Dice values, averaged across regions, were 0.81±0.02 using label propagation and fusion for the preterm population, and 0.81±0.02 using the single registration of a MPNA for the term population. Segmentations of 36 further unsegmented target images of developing brains yielded visibly high-quality results. This registration approach allows the rapid construction of automatically labeled age-specific brain atlases for neonates and the developing brain.

Introduction

Mapping the newborn brain is a significant challenge in neuroimaging. Unlike adult brains, neonatal brains undergo rapid growth and exhibit lower tissue contrast due to incomplete myelination. These factors make boundaries between anatomical structures difficult for standard algorithms to detect. Historically, the field has lacked "gold standard" neonatal data, often relying on adult templates that do not reflect infant brain proportions. To address this, we developed a robust framework for automatically segmenting neonatal brain MRIs into 50 distinct anatomical regions using a new manually segmented dataset.

Methods

The study relies on the ALBERTs (Average Landmark-Based Epitome of Real-Time Segments) cohort: 20 neonates (5 term-born, 15 preterm scanned at term-equivalent age) whose brains were manually segmented into 50 regions.

We compared two primary automated strategies:

  1. Multi-atlas Label Propagation and Fusion: Multiple manually labeled atlases are registered to a single target image, followed by a decision fusion (voting) process to determine final voxel labels.
  2. Probabilistic Templates (MPNA): A "Maximum Probability Neonatal ALBERT" (MPNA) template was created by assigning each voxel the most frequent label from the manually segmented cohort.
Figure 1: The iterative process for creating the average brain space (template) used for registration.

The registration pipeline utilized a three-step process—rigid, affine, and non-rigid B-spline transformations—specifically tuned for the low-contrast environment of neonatal MRI.

Results

Both automated methods achieved high accuracy, with a mean Dice overlap coefficient of 0.81 ± 0.02. This performance is considered excellent for complex subcortical structures in neonatal imaging.

Figure 6: Visual comparison of manual vs. automated segmentation (ALBERTs) for term and preterm brains across transverse, sagittal, and coronal planes.

The MPNA approach proved highly efficient, requiring only one registration per target brain (approx. 8 hours) compared to 20 registrations for the fusion approach (approx. 160 hours). For term-born infants, the MPNA approach actually outperformed the fusion method. Additionally, the system demonstrated robustness by successfully segmenting 36 "unseen" brains from preterm neonates scanned at birth (24–36 weeks gestational age), producing high-quality anatomical maps.

Discussion and Conclusion

This research demonstrates that the challenges of low-contrast neonatal imaging can be overcome using high-quality anatomical priors. Specificity in template selection is crucial; using "term" priors for "term" targets yielded the best results. While the current templates may struggle with extreme head shapes, the ALBERTs framework provides a scalable solution for large-scale studies of early brain development. Future work will explore group-wise registration to further reduce bias and improve accuracy in the most challenging anatomical regions.

Reference: Gousias IS, et al. (2013) Magnetic Resonance Imaging of the Newborn Brain: Automatic Segmentation of Brain Images into 50 Anatomical Regions. PLoS ONE 8(4): e59990.

Summary updated

Abstract

We studied methods for the automatic segmentation of neonatal and developing brain images into 50 anatomical regions, utilizing a new set of manually segmented magnetic resonance (MR) images from 5 term-born and 15 preterm infants imaged at term corrected age called ALBERTs. Two methods were compared: individual registrations with label propagation and fusion; and template based registration with propagation of a maximum probability neonatal ALBERT (MPNA). In both cases we evaluated the performance of different neonatal atlases and MPNA, and the approaches were compared with the manual segmentations by means of the Dice overlap coefficient. Dice values, averaged across regions, were 0.81±0.02 using label propagation and fusion for the preterm population, and 0.81±0.02 using the single registration of a MPNA for the term population. Segmentations of 36 further unsegmented target images of developing brains yielded visibly high-quality results. This registration approach allows the rapid construction of automatically labeled age-specific brain atlases for neonates and the developing brain.
BibTeX
@inproceedings{gousias2013magnetic,
  author = {Ioannis S. Gousias and Alexander Hammers and Serena J. Counsell and Latha Srinivasan and Mary Rutherford and Rolf A. Heckemann and Joseph V. Hajnal and Daniel Rueckert and A. David Edwards},
  title = {{Magnetic Resonance Imaging of the Newborn Brain: Automatic Segmentation of Brain Images into 50 Anatomical Regions}},
  booktitle = {PLoS},
  year = {2013},
  doi = {10.1371/journal.pone.0059990},
}
Figure