Hierarchical Clustering Diffusion Model for fMRI Functional Connectivity to Enhance Autism Spectrum Disorder Diagnosis

Objective: Functional magnetic resonance imaging (fMRI) data, particularly functional connectivity matrices, are crucial for studying brain disorders like Autism Spectrum Disorder (ASD). However, data scarcity often limits the performance of diagnostic models. We address this challenge by leveraging generative diffusion models for data augmentation. Approach: We introduce a novel transformer-based latent diffusion model, the Hierarchical Clustering Connective Diffusion Unit (HC-CDU), designed to synthesize realistic fMRI functional connectivity matrices. For comprehensive evaluation, we also analyze a non-hierarchical variant Connective Diffusion Unit (CDU) to quantify the specific impact of the integrated clustering module. Main Results: Our models effectively generate high-fidelity connectivity patterns, quantitatively confirmed by a notable reduction in Mean Absolute Error (MAE) values compared to baseline reconstruction, demonstrating an improvement of up to 3.61%. In classification tasks on the ABIDE-I dataset, the HC-CDU with x1 augmentation demonstrated a significant improvement, with AUC enhancing by up to 4.29% in specific folds over the no-augmentation baseline, showcasing enhanced discriminative power. While higher augmentation factors (x4) led to a decrease in overall accuracy and AUC for both variants, they notably increased Sensitivity, indicating a complex trade-off. Significance: This work underscores the substantial potential of the HC-CDU for augmenting complex fMRI data, offering a robust approach to mitigate data scarcity and improve diagnostic capabilities in neuroimaging studies for psychiatric disorders.