Geometric Deep Learning of the Human Connectome Project Multimodal Cortical Parcellation

Logan Z. J. Williams, Abdulah Fawaz, Matthew F. Glasser, A. David Edwards, Emma C. RobinsonLecture Notes in Computer Science 2021 · DOI · OpenAlex

Abstract

Understanding the topographic heterogeneity of cortical organisation is an essential step towards precision modelling of neuropsychiatric disorders. While many cortical parcellation schemes have been proposed, few attempt to model inter-subject variability. For those that do, most have been proposed for high-resolution research quality data, without exploration of how well they generalise to clinical quality scans. In this paper, we benchmark and ensemble four different geometric deep learning models on the task of learning the Human Connectome Project (HCP) multimodal cortical parcellation. We employ Monte Carlo dropout to investigate model uncertainty with a view to propagate these labels to new datasets. Models achieved an overall Dice overlap ratio of >0.85 ± 0.02. Regions with the highest mean and lowest variance included V1 and areas within the parietal lobe, and regions with the lowest mean and highest variance included areas within the medial frontal lobe, lateral occipital pole and insula. Qualitatively, our results suggest that more work is needed before geometric deep learning methods are capable of fully capturing atypical cortical topographies such as those seen in area 55b. However, information about topographic variability between participants was encoded in vertex-wise uncertainty maps, suggesting a potential avenue for projection of this multimodal parcellation to new datasets with limited functional MRI, such as the UK Biobank.

Introduction

The human cerebral cortex is a complex, folded sheet where functional regions vary significantly between individuals. "Cortical parcellation"—the process of dividing this surface into distinct areas based on structure, function, and connectivity—is crucial for neuroscientific research. While the Human Connectome Project (HCP) established a "gold standard" multimodal parcellation, the original method relied on a Multi-Layer Perceptron (MLP) that classified vertices independently. This approach often ignores spatial context and struggles with "atypical topographies," where specific regions shift, reshape, or fragment in certain individuals.

This study investigates whether Geometric Deep Learning (gDL), designed to operate directly on non-Euclidean surfaces like the brain's curved cortex, can better capture these spatial relationships and improve parcellation accuracy.

Methods

The researchers benchmarked four gDL architectures, categorized into spatial and spectral methods:

  • Spatial Methods (Spherical U-Net, MoUNet): These simulate traditional image filters by passing localized kernels over the brain's surface.
  • Spectral Methods (ChebNet, GConvNet): These utilize graph-based convolutions in the frequency domain of the brain's structure.

The cortex was modeled as an icosphere, allowing for U-Net-style pooling and upsampling. To quantify model confidence, the team implemented Monte Carlo dropout, generating uncertainty maps by running models multiple times with random neuron deactivation. The study used data from 390 HCP participants, incorporating 110 multimodal features per subject, including cortical thickness, myelin maps (T1w/T2w), and fMRI measures.

Key Results

The gDL models demonstrated high performance, with GConvNet emerging as the superior individual model. The authors suggest its success stems from a small kernel size that effectively captures fine-grained, localized features of the brain's "fingerprint." An ensemble approach, averaging predictions from all four models, further improved accuracy, indicating that spatial and spectral methods capture complementary information.

Performance metrics for different geometric deep learning models and their ensemble.

Quantitatively, the models achieved an overall Dice overlap ratio of >0.85 ± 0.02. Performance was strongest in primary sensory areas like the visual cortex (V1) but lower in regions with high biological variability, such as the medial frontal lobe and insula.

Discussion: Uncertainty and Atypical Topographies

A significant challenge remains in mapping "atypical topographies," such as the fragmented structure of Area 55b. While the gDL models tended to produce "typical," contiguous shapes—essentially smoothing over individual variations—the generated uncertainty maps effectively flagged these regions.

Comparison of ground truth labels (a) and model estimates (b) for the topologically variable Area 55b across different modalities.

This suggests that uncertainty estimation is not merely an error metric but a valuable tool for identifying unique individual biology. Future work will focus on refining loss functions beyond standard Dice overlap and testing these models on lower-quality clinical scans, such as the UK Biobank, to determine if structural data alone can support accurate parcellation in data-constrained environments.

Summary updated

BibTeX
@inproceedings{williams2021geometric,
  author = {Logan Z. J. Williams and Abdulah Fawaz and Matthew F. Glasser and A. David Edwards and Emma C. Robinson and Emma Claire Robinson and Logan Zane John Williams},
  title = {{Geometric Deep Learning of the Human Connectome Project Multimodal Cortical Parcellation}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2021},
  doi = {10.1007/978-3-030-87586-2\_11},
}

Loading related content...

Figure