Contrasformer: A Brain Network Contrastive Transformer for Neurodegenerative Condition Identification

TL;DR

Contrasformer tackles neurological disorder identification from fMRI brain networks by addressing sub-population distribution shifts and the lack of node-identity awareness. It introduces a contrast graph encoder with two-stream attention to produce disease-discriminative priors and a cross decoder with identity embedding to fuse these priors into node representations. The model is trained with four losses, including a novel ROI-level contrastive loss that enforces ROI identity consistency across subjects, plus sparsity and cluster constraints. Empirical results across four diverse datasets show consistent improvements over 13 baselines, with interpretable biomarker patterns aligning with neuroscience knowledge. This work enhances the applicability of Transformer-based models to dense brain graphs and provides a framework for robust, interpretable neurodiagnostics.

Abstract

Understanding neurological disorder is a fundamental problem in neuroscience, which often requires the analysis of brain networks derived from functional magnetic resonance imaging (fMRI) data. Despite the prevalence of Graph Neural Networks (GNNs) and Graph Transformers in various domains, applying them to brain networks faces challenges. Specifically, the datasets are severely impacted by the noises caused by distribution shifts across sub-populations and the neglect of node identities, both obstruct the identification of disease-specific patterns. To tackle these challenges, we propose Contrasformer, a novel contrastive brain network Transformer. It generates a prior-knowledge-enhanced contrast graph to address the distribution shifts across sub-populations by a two-stream attention mechanism. A cross attention with identity embedding highlights the identity of nodes, and three auxiliary losses ensure group consistency. Evaluated on 4 functional brain network datasets over 4 different diseases, Contrasformer outperforms the state-of-the-art methods for brain networks by achieving up to 10.8\% improvement in accuracy, which demonstrates its efficacy in neurological disorder identification. Case studies illustrate its interpretability, especially in the context of neuroscience. This paper provides a solution for analyzing brain networks, offering valuable insights into neurological disorders. Our code is available at \url{https://github.com/AngusMonroe/Contrasformer}.

Figures (7)

• Figure 1: The distribution of the original feature and Contrasformer representation for subjects from multiple sites and scanning duriation in ABIDE dataset. Each point in the figure represents a subject and different colors denote the sites these subjects are acquired from or their lengths of the BOLD signals. The representation of each subject is obtained by mean pooling and visualized by t-SNE van2008visualizing. Compared with (c) and (d), (a) and (b) exhibit obvious distribution shifts.
• Figure 2: The framework of Contrasformer for neurological disorder identification, using Autism as an example.
• Figure 3: The architecture of contrast graph encoder. Each group of brain networks is fed into the two-stream attention to obtain a summary graph. The contrast graph is generated by contrasting the summary graphs of different groups.
• Figure 4: The detail of two-stream attention. The ROI- and subject-wise attention blocks compute self-attention from different views of the input. Parameters of self-attention inside these two branches are independent.
• Figure 5: The architecture of the cross decoder. The generated contrast graph is incorporated with the identity-embedded brain network by a cross-attention for the downstream representation learning.