Abstract:Brain networks/graphs have been widely recognized as powerful and efficient tools for identifying neurological disorders. In recent years, various graph neural network models have been developed to automatically extract features from brain networks. However, a key limitation of these models is that the inputs, namely brain networks/graphs, are constructed using predefined statistical metrics (e.g., Pearson correlation) and are not learnable. The lack of learnability restricts the flexibility of these approaches. While statistically-specific brain networks can be highly effective in recognizing certain diseases, their performance may not exhibit robustness when applied to other types of brain disorders. To address this issue, we propose a novel module called Brain Structure Inference (termed BSI), which can be seamlessly integrated with multiple downstream tasks within a unified framework, enabling end-to-end training. It is highly flexible to learn the most beneficial underlying graph structures directly for specific downstream tasks. The proposed method achieves classification accuracies of 74.83% and 79.18% on two publicly available datasets, respectively. This suggests an improvement of at least 3% over the best-performing existing methods for both tasks. In addition to its excellent performance, the proposed method is highly interpretable, and the results are generally consistent with previous findings.

Learnable Brain Connectivity Structures for Identifying Neurological Disorders