Abstract
Mapping the functional regions of the cerebral cortex during the first two years of life is essential for understanding early brain maturation. However, the rapid changes in brain shape and size during this period make consistent longitudinal parcellation challenging. We present a multi-atlas based method for the simultaneous labeling of infant cortical surfaces across multiple time points. By employing an energy minimization framework that integrates local atlas data, geometry-aware spatial smoothness, and temporal consistency, our approach achieves superior accuracy and longitudinal stability compared to existing tools like FreeSurfer.
Introduction
The human brain undergoes explosive development during the first two years of life. Accurately parcellating the cerebral cortex into functional regions is critical for studying both healthy development and clinical conditions. Traditional parcellation methods often process each time point independently, leading to temporal "jitter" or inconsistent labels across scans. Alternatively, propagation-based methods can introduce bias toward a specific reference age.
To address these issues, we propose a simultaneous labeling framework that treats all longitudinal scans of a subject as a single optimization problem. This approach ensures that labels are consistent over time and unbiased toward any particular developmental stage.
Methods
The labeling task is formulated as an energy minimization problem solved using alpha-expansion graph cuts. The energy function consists of three primary components:
- Data Fitting Term ($E_d$): A multi-atlas approach is used, where 39 manually labeled expert atlases are adaptively weighted based on local cortical folding (mean curvature) similarity to the subject.
- Spatial Smoothness Term ($E_s$): This term is geometry-aware, allowing for label discontinuities at sulcal bottoms where functional boundaries naturally occur, thus preventing over-smoothing across distinct regions.
- Temporal Consistency Term ($E_t$): This term acts as a longitudinal constraint, encouraging vertices to maintain consistent labels across time points when local folding patterns remain stable.
The preprocessing pipeline includes rigid alignment of multi-modal images (T1, T2, FA), 4D level-set segmentation, and spherical mapping to establish vertex-wise correspondences across the longitudinal sequence.
Key Insights
A core innovation of this method is the use of sulcal bottoms as anchors. Recognizing that cytoarchitectural boundaries often align with the valleys of cortical folds, the algorithm respects these natural boundaries. Furthermore, by optimizing all time points simultaneously in a 4D graph, the model can leverage high-quality information from later scans (e.g., at 18 months) to resolve ambiguities in earlier, less developed scans (e.g., at birth).
Results
The proposed method was evaluated on a dataset of 13 healthy infants scanned at six intervals from birth to two years. Performance was compared against the longitudinal pipeline of FreeSurfer and a version of the proposed method without temporal constraints.
Quantitative results demonstrate that the proposed method achieves significantly higher accuracy, with average Dice coefficients of 0.941 for the precentral gyrus (PreCG) and 0.939 for the superior temporal gyrus (STG), compared to approximately 0.91 for FreeSurfer. Additionally, the method showed superior temporal stability, with a mean symmetric boundary distance of 0.52mm, nearly twice as consistent as FreeSurfer's 0.90mm. The entire optimization process is efficient, taking roughly 10 minutes per subject.
Conclusion
This simultaneous labeling framework provides a robust and unbiased tool for longitudinal infant brain studies. By effectively balancing local anatomical features with global temporal consistency, it offers a more accurate map of the growing brain, paving the way for deeper insights into early neurodevelopmental processes.
Summary updated