Conditional diffusion model with cross-attention that synthesizes subject-specific 3D intrinsic connectivity networks from resting-state fMRI.
No providers recorded yet. Browse all providers
Intrinsic connectivity networks (ICNs) — the spatially coherent sets of brain regions whose resting activity fluctuates together — are almost universally extracted with independent component analysis. ICA is fast, interpretable, and linear, and that last property is the constraint: it assumes the observed fMRI signal is a linear mixture of statistically independent sources, so nonlinear structure in the underlying networks has nowhere to go. Attempts to model that structure with generative networks have run into the usual failure modes, with progressive GANs producing artifacts and collapsing onto a narrow set of outputs.
CondAttn-DDPM applies denoising diffusion to the problem. Developed by Meenu Ajith and Vince Calhoun at the Tri-Institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS), a joint center of Georgia State University, the Georgia Institute of Technology, and Emory University, the work presents two frameworks. The ImageGen framework is an unconditional DDPM producing 2D and 3D ICN volumes; CondAttn-DDPM adds an attention mechanism and conditions the reverse diffusion process on a subject's own resting-state fMRI, so that generated networks reflect that individual rather than the group average.
The result is a per-subject decomposition obtained from a single frozen checkpoint. At inference the model receives only the rs-fMRI data and produces the corresponding ICNs, which places it in the same family as other narrow-domain conditional diffusion models in biology such as Lung-DDPM and scDiffusion — generative models trained once and then applied to new samples without refitting.
The denoising backbone is a U-Net with three downsampling stages, each a 2D or 3D convolution followed by max pooling, a convolutional bottleneck, and a symmetric upsampling path with skip connections; the conditional variant processes fMRI and ICN inputs jointly through an attention mechanism. Diffusion uses 1,000 discrete timesteps. Training ran for 20 epochs at batch size 32, checkpointing on best loss.
Data came from UK Biobank resting-state fMRI for 10,000 subjects, acquired on 3T Siemens Skyra scanners in an older adult cohort aged 53 to 87 (mean 69.75 ± 7.43 years). Preprocessing followed the standard FSL chain, ending in ICA-FIX artifact removal. Reference ICNs were estimated with NeuroMark, a fully automated spatially constrained ICA using the NeuroMark_fMRI_1.0 template of 53 components derived from HCP and Genomics Superstruct Project decompositions.
Although the abstract frames the comparison as qualitative, the paper reports quantitative agreement with ICA on the external HCP set: across 500 generated images per model, mean MS-SSIM was 0.72, MSE 0.001, and PSNR 27. The authors caution that closer agreement with ICA is not straightforwardly better, since the point of the diffusion model is to represent nonlinear structure a linear decomposition cannot. Against progressive GANs the comparison is clearer, with diffusion outputs free of the artifacts and mode collapse seen in the GAN baselines. For between-subject variability on ICN 3, CondAttn-DDPM averaged 0.82, 0.75, 0.81, and 0.82 across four subjects, against 0.74, 0.73, 0.72, and 0.71 for ICA.
The intended use is individualized functional network profiling: producing a complete set of ICNs for one person from one resting-state scan, in a form that supports comparison across sessions and across people. That output feeds precision-medicine workflows in psychiatry and neurology where group-average templates obscure the individual differences that matter, and it gives neuroimaging methods researchers a nonlinear alternative to ICA for studying how network topography varies across a cohort.
The work extends conditional diffusion into subject-level connectome synthesis and, unlike much of the generative neuroimaging literature, tests generalization on an independent cohort rather than a held-out split of the same one. Its limits are stated plainly by the authors: sampling 3D volumes through 1,000 denoising steps is computationally heavy and constrains scaling to larger datasets and higher resolutions, and output quality tracks the quality of the input scans. The preprint proposes replacing the U-Net with a transformer backbone and moving generation into a compressed latent space. It awaits peer review, carries a CC BY-NC-ND license, and no code or trained weights have been released.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.