TMN: Learning multi-timescale functional connectivity for identifying brain disorders.

BACKGROUND: Functional connectivity (FC) has been used to identify brain disorders. The present study aimed to identify brain disorders by FC across multiple timescales. METHODS: We first segmented the resting-state fMRI signals to construct multiple timescale functional connectivity (mFC) between brain regions. Next, we developed a deep multiple instance learning (MIL) approach, namely the Two-stage Multi-stream Network (TMN), to capture spatio-temporal patterns from the mFC. We evaluated the TMN in the ABIDE I dataset and the REST-Meta-MDD dataset. Furthermore, we proposed using the inputXgrad to explain the important features in the model. RESULTS: We achieved the best performance using the TMN model with mFC. Our findings indicated that mFC outperformed both static FC and the combination of static and dynamic FC in identification tasks. The model's explanation revealed that FC across all timescales contributed to the identification of brain disorders and highlighted the important FC that are strongly associated with these conditions. LIMITATIONS: The techniques used for data preprocessing can influence the model's performance, and this study requires further validation with a larger patient cohort and a broader range of brain disorders. CONCLUSIONS: The experimental results demonstrate that brain disorders can be effectively identified using the proposed TMN with mFC.