Cortex-Grounded Diffusion Models for Brain Image Generation

Synthetic neuroimaging data can mitigate critical limitations of real-world datasets, including the scarcity of rare phenotypes, domain shifts across scanners, and insufficient longitudinal coverage. However, existing generative models largely rely on weak conditioning signals, such as labels or text, which lack anatomical grounding and often produce biologically implausible outputs. To this end, we introduce Cor2Vox, a cortex-grounded generative framework for brain magnetic resonance image (MRI) synthesis that ties image generation to continuous structural priors of the cerebral cortex. It leverages high-resolution cortical surfaces to guide a 3D shape-to-image Brownian bridge diffusion process, enabling topologically faithful synthesis and precise control over underlying anatomies. To support the generation of new, realistic brain shapes, we developed a large-scale statistical shape model of cortical morphology derived from over 33,000 UK Biobank scans. We validated the fidelity of Cor2Vox based on traditional image quality metrics, advanced cortical surface reconstruction, and whole-brain segmentation quality, outperforming many baseline methods. Across three applications, namely (i) anatomically consistent synthesis, (ii) simulation of progressive gray matter atrophy, and (iii) harmonization of in-house frontotemporal dementia scans with public datasets, Cor2Vox preserved fine-grained cortical morphology at the sub-voxel level, exhibiting remarkable robustness to variations in cortical geometry and disease phenotype without retraining.

💡 Research Summary

Cor2Vox is a novel generative framework that produces anatomically faithful brain magnetic resonance images (MRIs) by conditioning the synthesis process on high‑resolution cortical surface information. The authors first construct a large‑scale statistical shape model of the cerebral cortex using over 33,000 UK Biobank scans. Each brain is represented by paired pial and white‑matter meshes (≈163 k vertices per hemisphere), which are converted into signed distance fields (SDFs). These SDFs serve as a continuous, spatially dense prior that captures the geometry of the cortex.

The core of Cor2Vox is a 3‑D Brownian Bridge Diffusion (BBD) model. Unlike conventional denoising diffusion probabilistic models (DDPMs) that start from pure Gaussian noise, BBD defines a stochastic bridge between two fixed endpoints: the MRI volume (source) and the cortical SDF (target). A linear schedule α_t = t/T interpolates between them, while a variance term δ_t = 2(α_t – α_t²) introduces stochasticity at the midpoint of the diffusion trajectory. The forward process analytically mixes the image and the SDF; the reverse process starts from the SDF and iteratively reconstructs the MRI.

To improve geometric consistency, four auxiliary conditions are supplied to the denoising network at every timestep: (1) the pial‑surface SDF, (2) the white‑matter‑surface SDF, (3) a binary edge map highlighting both surfaces, and (4) a cortical ribbon mask. These are concatenated with the noisy latent and fed into a 3‑D residual UNet equipped with convolutional and attention blocks. The network predicts the noise component, and the mean of the reverse transition is expressed as a linear combination of the current noisy volume, the target SDF, and the predicted noise, weighted by analytically derived coefficients. Training minimizes an ELBO‑derived loss that measures the KL divergence between the model’s reverse transition distribution and the analytically defined Brownian bridge.

Three application scenarios demonstrate the utility of Cor2Vox. First, anatomically consistent synthesis: given a cortical shape, the model generates MRIs that preserve fine sulcal patterns (e.g., central and superior temporal sulci) while varying scanner‑specific appearance. Quantitatively, Cor2Vox achieves a 30 % reduction in Fréchet Inception Distance (FID) and sub‑millimeter surface reconstruction error compared with a state‑of‑the‑art 3‑D DDPM. Second, progressive gray‑matter atrophy simulation: by traversing the latent space of the statistical shape model, the framework produces smooth, sub‑voxel cortical thinning trajectories that mimic disease progression in Alzheimer’s and frontotemporal dementia. Third, domain harmonization: frontotemporal dementia scans acquired in-house are transformed to match the appearance of ADNI data while retaining subject‑specific atrophy patterns. Whole‑brain segmentation Dice scores (SynthSeg+) improve by 0.02, and downstream classification AUC rises by 3 %.

Limitations include high memory consumption due to dense 3‑D SDFs and the UNet, restricting training to 1 mm³ resolution volumes. The current conditioning focuses solely on the cortex, leaving subcortical structures uncontrolled. Future work aims to incorporate multi‑tissue masks, multi‑scale diffusion, and memory‑efficient architectures to enable whole‑brain, high‑resolution synthesis.

In summary, Cor2Vox introduces a cortex‑grounded diffusion paradigm that bridges anatomical shape priors and image synthesis, delivering high‑quality, anatomically plausible brain MRIs without the need for large labeled datasets. Its ability to generate novel cortical geometries, simulate disease‑related atrophy, and harmonize cross‑site data makes it a valuable tool for neuroimaging research, data augmentation, and virtual cohort generation.