Longitudinal Cortical Registration for Developing Neonates

Hui Xue, Latha Srinivasan, Shuzhou Jiang, Mary A. Rutherford, A. David Edwards, Daniel Rueckert, Joseph V. HajnalMICCAI 2007 · DOI

Abstract

Tracking the rapid structural evolution of the neonatal brain is critical for identifying early neurodevelopmental markers. This paper presents an automated two-stage algorithm for longitudinal cortical registration. By adaptively smoothing mature, highly-folded surfaces to match the simpler geometry of earlier developmental stages, the method overcomes the limitations of standard non-rigid registration. Evaluation on a longitudinal dataset of 15 neonatal MR scans demonstrates significant improvements in registration accuracy, particularly across large developmental gaps, achieving a True Positive rate of 0.72 compared to 0.41 for standard methods.

Introduction

The human brain undergoes an explosion of growth and folding (gyrification) during the third trimester of pregnancy. For researchers studying preterm infants, tracking these changes is vital for identifying early markers of neurodevelopmental abnormalities. However, registering cortical surfaces across different gestational ages is notoriously difficult. A brain at 27 weeks is relatively smooth, while the same brain at 41 weeks is a complex labyrinth of folds. Traditional registration methods designed for adults often fail because they get "lost" in the complex folds of the more mature brain, failing to find correct anatomical correspondences in the younger, smoother brain.

Methods

The proposed pipeline avoids explicit spherical mapping, focusing instead on establishing point correspondences directly on the original surfaces.

1. Segmentation & Reconstruction

Because neonatal MRI has inverted contrast (white matter is brighter than gray matter on T2-weighted images), the team used a modified Expectation-Maximization (EM) algorithm with Markov Random Fields to accurately segment the tissues. Surfaces were then reconstructed using a level-set approach.

Segmentation and reconstruction of the neonatal cortex. (a) Original T2-weighted MRI. (b-d) Comparison of segmentation results highlighting the accuracy of the reconstructed cortical boundary.

2. Global Alignment

Instead of using the common Iterative Closest Point (ICP) method, which is prone to local minima, voxel-based image registration was used to calculate an initial affine transformation.

3. Adaptive Surface Relaxation

To bridge the gap between a smooth young brain and a folded older brain, the authors "relax" (inflate/smooth) the more mature surface. They use two dimensionless metrics—the Intrinsic Curvature Index (ICI) and Mean Curvature $L^2$ norm (MLN)—to determine when the mature brain has been smoothed enough to match the complexity of the younger brain.

4. Non-rigid Registration

Once the surfaces are "complexity-matched," a non-rigid registration based on Free-Form Deformations (FFDs) is applied. This uses B-splines to smoothly deform one surface to match the other, minimizing the average symmetric spatial distance between them.

The proposed longitudinal registration pipeline. The mature folded brain (a) is inflated (1) to match the complexity of the younger brain (c). After registration (2), the surface is deformed (3) to establish point correspondences, resulting in the final rendered mapping (e).

Key Insights

  • Smoothing as Pre-processing: Reducing the complexity of a target image can improve registration accuracy. By "un-folding" the 41-week brain until it resembles a 27-week brain, the algorithm more easily identifies primary anatomical landmarks.
  • Avoiding Local Optima: High-frequency folding patterns act as local minima for registration algorithms. Partial inflation acts as a low-pass filter, allowing the algorithm to focus on global anatomical correspondence before matching fine details.
  • Direct Surface Correspondence: Working on original polygon meshes rather than spherical proxies preserves the ability to describe local evolution in 3D space.

Results

The method was evaluated by propagating manual sulcus labels from earlier scans to later scans. The proposed method (NR+I: Non-rigid + Inflation) consistently outperformed standard non-rigid registration (NR) and intensity-based registration (IM).

  • Long-range Accuracy: In mapping from the 1st scan (smooth) to the 3rd scan (highly folded), the NR+I method achieved a True Positive (TP) rate of 0.72, whereas standard non-rigid registration only managed 0.41.
  • Short-range Precision: For the 1st to 2nd scan mapping, the proposed method reached a near-perfect 0.97 TP rate with a low 0.12 False Positive rate.

Conclusion

While inflation helps align major landmarks, it can smooth out secondary sulci. Future work will incorporate sulcus shape into similarity measures to protect smaller features and apply this technique to fetal brain scans to create a comprehensive atlas of normal cortical growth.

Summary updated

BibTeX
@inproceedings{xue2007longitudinal,
  author = {Hui Xue and Latha Srinivasan and Shuzhou Jiang and Mary A. Rutherford and A. David Edwards and Daniel Rueckert and Joseph V. Hajnal},
  title = {{Longitudinal Cortical Registration for Developing Neonates}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2007},
  doi = {10.1007/978-3-540-75759-7\_16},
}
Figure