Explainable Self-Supervised Dynamic Neuroimaging Using Time Reversal

Functional magnetic resonance imaging data pose significant challenges due to their inherently noisy and complex nature, making traditional statistical models less effective in capturing predictive features. While deep learning models offer superior performance through their non-linear capabilities, they often lack transparency, reducing trust in their predictions. This study introduces the Time Reversal (TR) pretraining method to address these challenges. TR aims to learn temporal dependencies in data, leveraging large datasets for pretraining and applying this knowledge to improve schizophrenia classification on smaller datasets.

We pretrained an LSTM-based model with attention using the TR approach, focusing on learning the direction of time in fMRI data, achieving over 98 % accuracy on HCP and UK Biobank datasets. For downstream schizophrenia classification, TR-pretrained weights were transferred to models evaluated on FBIRN, COBRE, and B-SNIP datasets. Saliency maps were generated using Integrated Gradients (IG) to provide post hoc explanations for pretraining, while Earth Mover's Distance (EMD) quantified the temporal dynamics of salient features in the downstream tasks.

TR pretraining significantly improved schizophrenia classification performance across all datasets: median AUC scores increased from 0.7958 to 0.8359 (FBIRN), 0.6825 to 0.7778 (COBRE), and 0.6341 to 0.7224 (B-SNIP). The saliency maps revealed more concentrated and biologically meaningful salient features along the time axis, aligning with the episodic nature of schizophrenia. TR consistently outperformed baseline pretraining methods, including OCP and PCL, in terms of AUC, balanced accuracy, and robustness.

This study demonstrates the dual benefits of the TR method: enhanced predictive performance and improved interpretability. By aligning model predictions with meaningful temporal patterns in brain activity, TR bridges the gap between deep learning and clinical relevance. These findings emphasize the potential of explainable AI tools for aiding clinicians in diagnostics and treatment planning, especially in conditions characterized by disrupted temporal dynamics.