SpaRG: Sparsely Reconstructed Graphs for Generalizable fMRI Analysis.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Deep learning models analyze resting-state functional Magnetic Resonance Imaging (rs-fMRI) data to find patterns related to psychiatric disorders and traits.
• Interpreting deep learning findings in fMRI is challenging due to data sensitivity to scanning effects and visualization difficulties.

Purpose of the Study:

• To propose a novel approach combining sparsification and self-supervision to improve the interpretability and robustness of deep learning models in fMRI analysis.
• To develop a method that identifies and utilizes only the most informative brain connections for classification tasks, reducing noise and improving generalization.

Main Methods:

• A joint end-to-end training framework involving a sparse input mask, a variational autoencoder (VAE), and a downstream classifier.
• Optimization of the sparse mask and VAE using unlabeled data from additional sites to retain generalizable input features.
• Evaluation of the Sparsely Reconstructed Graphs (SpaRG) method on the ABIDE dataset for sex classification, including adaptation to out-of-distribution sites.

Main Results:

• SpaRG effectively identifies a small subset (1%) of highly informative connections for classification tasks, even with a coarse brain parcellation (64 regions).
• The method demonstrates improved classification accuracy across domains, including adaptation to new, unseen datasets.
• SpaRG enhances the interpretability of deep learning models by focusing on essential functional connections.

Conclusions:

• The proposed sparsification and self-supervision approach (SpaRG) offers a robust and interpretable method for analyzing fMRI data with deep learning.
• SpaRG successfully mitigates challenges in fMRI data analysis, leading to improved performance and generalizability in classification tasks.
• This method provides a valuable tool for uncovering meaningful patterns in brain connectivity for psychiatric and trait-related research.