Abstract
Graph Neural Networks (GNNs) have emerged as powerful tools for brain network analysis, yet their effectiveness is often hindered by the scarcity of labeled medical data and the potential for "anatomical distortion" introduced by standard data augmentation techniques. This paper introduces Self-Promoted Clustering-based Contrastive Learning (SPCCL), a novel pre-training framework designed to leverage unlabeled healthy subject data without compromising biological integrity. By employing a self-promoted clustering mechanism that prioritizes challenging samples, SPCCL achieves state-of-the-art accuracy in identifying schizophrenia and provides interpretable insights into the disorder's neural correlates.
Introduction
Brain networks, where nodes represent anatomical regions and edges represent structural or functional connections, provide a window into the complex organization of the human brain. While GNNs are adept at modeling these graphs, they typically require large datasets that are difficult to obtain in clinical settings. To address this, researchers often turn to Graph Contrastive Learning (GCL). However, standard GCL relies on random edge or node perturbations. In neuroimaging, such augmentations are problematic because edges represent physical white matter tracts or specific functional correlations; random deletions can "damage" the underlying biological topology and lead to neurologically misleading representations.
Methodology
The SPCCL framework avoids "illegal" augmentations by focusing on the inherent structure of brain networks and utilizing a supervised contrastive approach on pre-training data. The architecture is composed of four primary modules:
- Siamese GCN Module: Extracts node-level embeddings from pairs of structural and functional brain networks using shared weights.
- Self-Supervised Readout: Employs a Bi-directional LSTM and an Attention mechanism to aggregate node embeddings into a graph-level vector, effectively handling node ordering bias and highlighting discriminative features.
- Contrastive Clustering Module: Uses a supervised contrastive loss to pull samples from the same group (e.g., healthy controls) together while pushing different groups apart in the embedding space.
- Self-Promoted Contrastive Pairs: An active learning component that identifies "hard" samples—those furthest from their cluster centers—and promotes them for the next training round, forcing the model to learn more nuanced class boundaries.
Experimental Results
The model was pre-trained on a dataset of 70 healthy adults and then fine-tuned on a schizophrenic dataset (27 patients, 27 healthy controls). SPCCL was benchmarked against several baselines, including Support Vector Classifiers (SVC) and the state-of-the-art GCBN method.
- Classification Performance: SPCCL achieved a significant accuracy of 85.2%, outperforming GCBN's 73.1%. It also maintained a high F1 score of 0.858, demonstrating a robust balance between precision and recall.
- Biological Interpretability: A key advantage of SPCCL is its interpretability. By analyzing the attention weights, the researchers identified the top 20 most influential brain regions for classification. These regions—including the temporal gyrus, frontal gyrus, insula, and amygdala—align closely with established neuroscience literature regarding schizophrenia.
Discussion and Conclusion
SPCCL demonstrates that augmentation-free contrastive learning is not only viable but superior for biological data where structural integrity is paramount. By pre-training on readily available healthy data and focusing on "hard" samples through self-promotion, the model learns a robust baseline of brain organization.
Future work will involve validating the model on larger, multi-site datasets to ensure generalizability and extending the framework to multi-class diagnostic problems or longitudinal disease progression forecasting. Despite the added computational complexity of the iterative clustering process, the gains in accuracy and biological plausibility make SPCCL a promising direction for neuroimaging research.
Summary updated