2026 ECNP Congress

PS01-1148

MIND-based morphometric similarity networks for bipolar disorder classification via graph neural networks

Inês Sampaio — G. Poli, A. Pigoni, M. Bellani, F. Benedetti, I. Nenadić, M. L. Philips, F. Piras, J. C. Soares, Y. Torrente, L. N. Yatham, A.M. Bianchi, E. Maggioni, P. Brambilla · Fondazione IRCCS Ca’ Granda Ospedale Maggiore Policlinico, Department of Neurosciences and Mental Health, Milano, Italy

2026 ECNP Congress

Introduction

Morphometric similarity networks (MSN) have demonstrated considerable potential as proxies for the structural brain connectome [1], effectively capturing the brain architectome when derived through advanced algorithms such as Morphometric Inverse Divergence (MIND) [2]. This framework is of particular relevance to psychiatric conditions in which neurodevelopmental or neuroprogressive pathophysiological mechanisms are implicated, including bipolar disorder (BD). However, the utility of MSN, and of MIND networks specifically, has not yet been investigated in this context, representing a gap in the literature that the present study aims to address.

Aims

The primary aim of this study was to investigate the utility of MIND-derived morphometric similarity networks (MIND networks) for the classification of BD relative to healthy controls (HC) in the multi-site StratiBip cohort, and to evaluate the impact of different brain parcellation schemes on graph topology and classification performance.

Methods

T1-weighted structural MRI scans were acquired across seven sites adhering to the StratiBip consortium using 3T scanners, yielding an age-balanced sample of N=997 subjects, partitioned into a training set (n_HC=422 ; n_BD=351), and an independent test set (n_HC=90 ; n_BD=53). Images were preprocessed using the CAT12 toolbox within SPM. MIND networks were constructed from voxel-based gray matter volume (GMV) maps, with pairwise morphometric similarity computed between regions of interest defined according to three brain atlas parcellations: NatBrainLab, Neuromorphometrics, and Brodmann. MIND similarity values were harmonized using ComBat [3], with site-correction parameters estimated exclusively in the training set to preserve the independence of the test set[4]. Graph convolutional networks (GCN)[5] were employed to probe BD classification performance across varying graph topologies, defined by: (i) three brain parcellation schemes, yielding different node and edge configurations; and (ii) multiple node feature combinations, including MIND similarity values, GMV, and demographic variables (age and sex). Graph sparsification was applied using proportional thresholding with a top-k approach, where k=30% surviving connections.

Results

Harmonization quality was assessed through two-dimensional UMAP projections, confirming that site-associated clustering was no longer apparent following harmonization. GCN classification performance was highest for node feature combinations incorporating MIND similarity values with age (MIND + age) and MIND similarity values with age, sex, and GMV (MIND + age + sex + GMV), both using the Neuromorphometrics parcellation, achieving a best classification accuracy of 66.27%. The Neuromorphometrics atlas consistently outperformed the NatBrainLab and Brodmann parcellations across all remaining node feature combinations examined.

Conclusions

This study demonstrates the potential of morphometric similarity networks for investigating BD and, more broadly, psychiatric conditions characterized by neurodevelopmental or neuroprogressive alterations. The graph neural network framework provided an effective means of analysing graph-structured brain data and advanced the predictive modelling of BD. Notably, the integration of age with MIND similarity values consistently emerged as a relevant feature combination, suggesting a meaningful interaction between morphometric network architecture and neurodevelopmental or aging-related trajectories in BD.

References

  1. I. Sebenius, L. Dorfschmidt, J. Seidlitz, A. Alexander-Bloch, S. E. Morgan, and E. Bullmore, “Structural MRI of brain similarity networks,” Nat. Rev. Neurosci., vol. 26, no. 1, pp. 42–59, 2025, doi: 10.1038/s41583-024-00882-2.
  2. I. Sebenius et al., “Robust estimation of cortical similarity networks from brain MRI,” Nat. Neurosci., vol. 26, no. 8, pp. 1461–1471, 2023, doi: 10.1038/s41593-023-01376-7.
  3. W. E. Johnson, C. Li, and A. Rabinovic, “Adjusting batch effects in microarray expression data using empirical Bayes methods,” Biostatistics, vol. 8, no. 1, pp. 118–127, 2007, doi: 10.1093/biostatistics/kxj037.
  4. I. W. Sampaio et al., “Comparison of Multi-site Neuroimaging Data Harmonization Techniques for Machine Learning Applications,” EUROCON 2023 - 20th Int. Conf. Smart Technol. Proc., pp. 307–312, 2023, doi: 10.1109/EUROCON56442.2023.10198911.
  5. T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” 5th Int. Conf. Learn. Represent. ICLR 2017 - Conf. Track Proc., pp. 1–14, 2017.