Hippocampal Segmentation From Longitudinal Infant Brain MR Images via Classification-Guided Boundary Regression

Yeqin Shao, Jaeil Kim, Yaozong Gao, Qian Wang, Weili Lin, Dinggang ShenIEEE Access 2019 · DOI

Introduction

The first year of life is a period of unprecedented brain development. Among the most critical structures is the hippocampus, a region central to memory formation and learning. Accurately mapping its growth is essential for identifying early biomarkers of neurodevelopmental disorders. However, segmenting the hippocampus in infant brain MRIs is notoriously difficult due to the unique physiological characteristics of the developing brain.

The Challenge of Infant Neuroimaging

Standard segmentation tools often fail when applied to infant scans. Unlike adult brains, infant MRIs exhibit extremely low contrast between gray and white matter, a result of ongoing myelination. Furthermore, the hippocampus is a small, complex structure that undergoes rapid morphological changes during the first 12 months.

Traditional "atlas-based" methods, which rely on warping a standard template, struggle with the high anatomical variability of infants. Similarly, deformable models often fail because they require distinct edges, which are frequently absent or blurred in early infant scans.

Proposed Methodology: Classification-Guided Boundary Regression

To address these challenges, researchers have introduced a multi-stage framework centered on Classification-Guided Boundary Regression (CBR). This approach shifts the focus from simple voxel labeling to predicting the distance to structural boundaries.

1. Multi-Modal and Longitudinal Context

The framework leverages multiple imaging modalities to compensate for low contrast. By combining T1-weighted and T2-weighted images with Fractional Anisotropy (FA) from Diffusion Tensor Imaging (DTI), the model gains structural clues that are invisible in standard anatomical scans.

Multi-modal infant brain images (T1, T2, and FA) showing the low contrast and the hippocampal region of interest.

Furthermore, by analyzing scans from the same subject at multiple time points (e.g., 2 weeks, 3, 6, 9, and 12 months), the algorithm can "look ahead." Since 12-month scans have higher contrast, they act as a temporal anchor, helping to resolve ambiguities in the earlier, noisier scans.

2. Boundary Regression and Refinement

Instead of a binary classification (deciding if a voxel is "in" or "out"), the system uses a Random Forest to predict a 3D displacement vector to the nearest hippocampal boundary. This regression is "guided" by local shape labels, providing the model with a better understanding of the overall geometry. To improve accuracy, the framework employs auto-context, an iterative process where the output of one forest serves as the input for the next, progressively sharpening the boundary estimation.

3. Deformable Model Integration

The final stage of the pipeline involves an Active Shape Model (ASM). Rather than searching the raw MRI data, the ASM is "pulled" toward the high-probability boundaries generated by the regression forest. This ensures that the final segmentation is both anatomically plausible and precise.

The transition from a refined boundary map to the final segmentation using a deformable model.

Results and Performance

The CBR framework was evaluated on a longitudinal dataset of 10 infants. It significantly outperformed state-of-the-art multi-atlas label fusion techniques:

  • Dice Similarity Coefficient (DSC): The method achieved an overlap of 71.12% at the challenging 2-week mark, rising to 79.87% by 12 months.
  • Average Surface Distance (ASD): Boundary placement error was reduced to just 0.63mm at 12 months.

Conclusion and Future Directions

By combining boundary regression with longitudinal and multi-modal context, this framework provides a robust solution for infant hippocampal segmentation. While currently limited by a small sample size that precludes deep learning, the authors plan to transition these techniques into neural architectures as larger datasets become available, potentially streamlining clinical neurodevelopmental assessments.

Summary updated

BibTeX
@article{shao2019hippocampal,
  author = {Yeqin Shao and Jaeil Kim and Yaozong Gao and Qian Wang and Weili Lin and Dinggang Shen},
  title = {{Hippocampal Segmentation From Longitudinal Infant Brain MR Images via Classification-Guided Boundary Regression}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2904143},
}
Figure