Brain Parcellation on the Cortical Mesh via Deep Modularity Network

Loading...
Thumbnail Image

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Cortical parcellation seeks to divide the cerebral cortex into a set of spatially contiguous parcels such that vertices within each parcel exhibit coherent resting-state functional connectivity (RSFC) patterns. Such parcels serve as candidate neurobiological units for studying large-scale brain organization, summarizing high-dimensional functional connectivity data, and supporting downstream analyses in network neuroscience \cite{Schaefer2018,Gordon2016,Laumann2015}. This problem is inherently local-global: biologically meaningful boundaries often coincide with sharp local transitions in RSFC across the cortical surface, while the interior of a cortical area is characterized by globally similar functional time courses. Classical parcellation methods therefore combine (i) a global similarity term that rewards grouping vertices with similar RSFC profiles and (ii) a local boundary term that discourages merging across strong RSFC gradients. A representative approach is the gradient-weighted Markov Random Field (gwMRF), which models unit-normalized time series using a von Mises--Fisher (vMF) mixture likelihood and adds a gradient-weighted pairwise prior that reduces boundary penalties where RSFC gradients are high. To ensure parcels remain spatially connected, gwMRF further introduces a spatial prior on spherical coordinates and relies on a staged optimization schedule in which the spatial constraint is gradually relaxed to avoid unrealistically round parcels \cite{Schaefer2018}. While effective, gwMRF requires repeated graph-cut inference within a coordinate-descent loop, is sensitive to cost scaling and initialization, and can become computationally heavy at high resolutions.

In this paper, we investigate whether a fully differentiable, modularity-based deep clustering approach can serve as a practical alternative to MRF-based parcellation on the cortical mesh. We propose a deep modularity parcellation model that constructs a gradient-weighted functional graph from RSFC gradient edge-density information, encodes per-vertex signals using a graph neural network (GNN), and learns soft cluster assignments by optimizing a modularity-derived objective (Deep Modularity Networks; DMoN). To bias solutions toward plausible parcels, we augment the modularity objective with a Laplacian smoothness regularizer defined on the cortical mesh, which promotes local label consistency and smooth boundaries, and we include a collapse regularizer to prevent degenerate clusterings. Using Human Connectome Project (HCP) rfMRI data, our preliminary experiments suggest that the proposed objective exhibits stable descent with faster convergence than gwMRF under matched cluster counts. We provide an end-to-end pipeline from RSFC gradient construction and graph formation to clustering and evaluation.

Description

Keywords

Brain parcellation , Graph clustering , Deep Modularity

Citation

Endorsement

Review

Supplemented By

Referenced By