Multiple Cortical Surface Correspondence Using Pairwise Shape Similarity

Pahal Dalal, Feng Shi, Dinggang Shen, Song WangMICCAI 2010 · DOI

Abstract

Establishing accurate anatomical correspondence across diverse human cortical surfaces is a fundamental challenge in neuroimaging. This paper presents a novel framework that organizes a population of brain surfaces into a "low-height tree" based on pairwise shape similarity. By strictly limiting the depth of the tree, the proposed method significantly reduces the accumulation of registration errors that typically plague traditional template-based or minimum spanning tree (MST) approaches. Experimental results on the LPBA40 dataset demonstrate that this method achieves higher accuracy and consistency while maintaining a significantly lower memory footprint than groupwise optimization techniques.

Introduction

Mapping the highly convoluted surface of the human cerebral cortex is essential for comparing anatomical features like cortical thickness or folding patterns across populations. Such comparisons are vital for understanding neurodevelopment and diagnosing neurological disorders. However, the immense structural variability between individuals makes establishing "correspondence"—ensuring a point on one brain represents the same landmark on another—extremely difficult.

Traditional approaches generally follow two paradigms: fixed template matching and groupwise optimization. Fixed templates compare all subjects to a single "average" brain, which often fails for subjects with atypical structures. Groupwise optimization attempts to align all subjects simultaneously; while potentially more accurate, it is computationally intensive and prone to sub-optimal local minima. Recent efforts using Minimum Spanning Trees (MST) to link similar brains have suffered from "tall" tree structures, where small registration errors compound as they propagate down long chains of subjects.

Methodology

The authors propose a Shape-Organization Tree that balances pairwise similarity with tree "shallowness" to limit error propagation.

  1. Similarity Metric: To determine brain similarity, the method utilizes an initial registration followed by the calculation of a dissimilarity score. This score is based on the alignment of gyri (ridges) and sulci (valleys). By using a binary representation of these features rather than raw curvature values, the metric becomes more robust to noise.
  2. Tree Construction: The process begins with a fully connected graph of all subjects, weighted by dissimilarity. The graph is pruned to its most essential connections, and a central subject is selected as the root. The tree is then built layer-by-layer to ensure a low height. This ensures that even structurally unique brains remain within a few steps of the root, minimizing the "telephone game" effect of error accumulation.
Comparison of tree topologies for population registration: (a) the proposed low-height tree, (b) a star graph (template-based), and (c) a standard spanning tree.
  1. Landmark Propagation: Once the tree is established, anatomical landmarks are sampled on the root brain and propagated through the branches to all other subjects in the population.

Experimental Evaluation

The proposed method was evaluated using the LONI Probabilistic Brain Atlas (LPBA40), which contains 40 MR images with 56 manually labeled anatomical structures. These labels serve as a ground truth to quantitatively verify if the mathematical mapping correctly identifies biological structures.

Results and Discussion

The low-height tree method outperformed standard approaches across several metrics:

  • Accuracy: The method achieved a total average error ($\Delta$) of 0.479, compared to 0.509 for fixed templates, 0.532 for MST-based methods, and 0.791 for groupwise optimization.
  • Consistency: Qualitative analysis using spherical maps showed that the proposed method correctly identified anatomical labels across 77.2% of the surface, the highest among all tested methods.
Qualitative consistency maps showing the percentage of correctly identified anatomical labels across the cortical surface for different registration methods.
  • Efficiency: The tree-based approach required less than 2GB of RAM, a significant improvement over the 11GB required by groupwise optimization for the same dataset.

Conclusion

By prioritizing both shape similarity and tree shallowness, the proposed low-height tree framework provides a robust and efficient solution for multiple cortical surface correspondence. While the initial $O(M^2)$ cost of calculating pairwise dissimilarities is a limitation for very large datasets, the method's accuracy and memory efficiency make it a powerful tool for standard neuroimaging cohorts. Future work will focus on optimizing the similarity calculations to scale the approach to larger populations.

Summary updated

BibTeX
@inproceedings{dalal2010multiple,
  author = {Pahal Dalal and Feng Shi and Dinggang Shen and Song Wang},
  title = {{Multiple Cortical Surface Correspondence Using Pairwise Shape Similarity}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2010},
  doi = {10.1007/978-3-642-15705-9\_43},
}
Figure