Explainable Self-Supervised Dynamic Neuroimaging Using Time Reversal

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Functional magnetic resonance imaging (fMRI) data present challenges for traditional models due to noise and complexity.
• Deep learning models offer improved performance but often lack transparency, hindering clinical trust.
• Existing methods struggle to capture temporal dynamics crucial for understanding neurological conditions.

Purpose of the Study:

• Introduce the Time Reversal (TR) pretraining method to enhance deep learning for fMRI analysis.
• Improve the accuracy and interpretability of schizophrenia classification using fMRI data.
• Leverage large datasets for pretraining to boost performance on smaller, specialized datasets.

Main Methods:

• Pretrained a Long Short-Term Memory (LSTM) network with attention using the TR approach on large fMRI datasets.
• Transferred TR-pretrained weights to models for schizophrenia classification on FBIRN, COBRE, and B-SNIP datasets.
• Utilized Integrated Gradients (IG) for saliency mapping and Earth Mover's Distance (EMD) for temporal dynamics analysis.

Main Results:

• TR pretraining significantly improved schizophrenia classification performance across all tested datasets (e.g., median AUC increased from 0.7958 to 0.8359 on FBIRN).
• Saliency maps showed more biologically relevant temporal features, aligning with schizophrenia's episodic nature.
• TR outperformed baseline pretraining methods (OCP, PCL) in AUC, balanced accuracy, and robustness.

Conclusions:

• The TR method enhances both predictive performance and interpretability in fMRI-based deep learning models.
• TR aligns model predictions with meaningful temporal brain activity patterns, increasing clinical relevance.
• Explainable AI tools like TR show promise for clinical diagnostics and treatment planning in conditions with disrupted temporal dynamics.