Conditional Denoising Diffusion Probabilistic Models with Attention for Subject-Specific Brain Network Synthesis

Abstract

The development of diffusion models, such as Glide, DALLE 2, Imagen, and Stable Diffusion, marks a significant advancement in generative AI for image synthesis. In this paper, we introduce a novel framework for synthesizing intrinsic connectivity networks (ICNs) by utilizing the nonlinear capabilities of denoising diffusion probabilistic models (DDPMs). This approach builds upon and extends traditional linear methods, such as independent component analysis (ICA), which are commonly used in neuroimaging studies. A central contribution of our work is the integration of attention mechanisms into conditional DDPMs, enabling the generation of subject-specific 3D ICNs. Conditioning the resting-state fMRI (rs-fMRI) data on the corresponding ICNs enables the extraction of individualized brain connectivity patterns, effectively capturing within-subject and between-subject variability. Unlike prior models limited to 2D visualization, this framework generates 3D representations, providing a more comprehensive depiction of ICNs. The models performance is validated on an external dataset to prevent over-fitting and for overall generalizability. Furthermore, comparative evaluations also demonstrate that the proposed DDPM-based approach outperforms state-of-the-art generative models in producing more detailed and accurate ICNs, as validated through qualitative assessments.