Denoising Diffusion Probabilistic Models for High-Fidelity fMRI Intrinsic Connectivity Network Data Generation

The emergence of diffusion models such as Glide, Dalle-2, Imagen, and Stable Diffusion marks a significant breakthrough in generative AI-based image generation. This paper introduces an rs-fMRI image synthesis framework that leverages the nonlinear capabilities of denoising diffusion probabilistic models (DDPMs) to overcome the limitations of linear methods like independent component analysis (ICA) in neuroimaging analysis. Unlike ICA, which assumes linearity, DDPMs capture the intricate and complex patterns inherent in neuroimaging data. Our approach advances from 2D to 3D representations, providing a comprehensive visualization of intrinsic connectivity networks (ICNs). This framework also addresses the challenge of sparse training datasets commonly encountered in deep learning applications for neuroimaging. Trained on a large database, our model captures the intricate variability of different ICNs, generating realistic connectivity patterns. The proposed method is evaluated quantitatively to compare the synthesized ICNs against ground truth data. Results demonstrate that the proposed DDPM-based framework showcases competitive performance accuracy in reflecting the true complexity of neural connectivity patterns.