Cortex-Grounded Diffusion Models for Brain Image Generation

TL;DR

Cor2Vox addresses the challenge of biologically plausible 3D brain MRI synthesis by grounding diffusion-based generation in cortical surface geometry. It introduces a cortex-conditioned Brownian bridge diffusion process that maps a cortex signed distance field $ ext{S}_c$ to MRI volumes, guided by auxiliary cortex representations and a large PCA-based cortical shape model derived from over $33{,}000$ UKB scans. The method yields high geometric fidelity, enables sub-voxel cortical atrophy simulations, and supports cross-dataset harmonization without retraining, outperforming multiple baselines on both image fidelity and cortical surface consistency. This cortex-grounded framework has significant potential for data augmentation, disease progression modeling, and robust cross-site analyses in neuroimaging.

Abstract

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.

Figures (10)

• Figure 1: Synthetic MRIs generated by Cor2Vox (top) and 3D-DDPM (bottom), with real MRIs shown for reference (center). Cortical surfaces were reconstructed from the MRIs using Vox2Cortex-Flow. Red lines indicate the central sulcus and superior temporal sulcus, two prominent cortical grooves in human brain anatomy. While Cor2Vox reliably preserves these sulci, they are often absent or significantly distorted in surfaces reconstructed from MRIs generated by 3D-DDPM. Such anatomical irregularities are virtually impossible to detect in conventional 2D slice-based views and require 3D surface reconstructions for reliable assessment. $\phi\in(0,1)$ denotes the spherical interpolation factor in the Cor2Vox shape model.
• Figure 2: Cor2Vox overview. a, In dependence on the application, input cortical surfaces can be created based on a statistical shape model of the cerebral cortex, surface-based simulations, or existing shapes from external databases. b, The cortical surface meshes, i.e., pial ($\mathcal{S}_p$) and white matter ($\mathcal{S}_w$) surfaces, are converted into a joint cortex signed distance field (SDF), $\mathcal{S}_c$. c, Cor2Vox leverages a shape-to-image Brownian bridge diffusion process to learn a stochastic mapping $f_\theta$ from the spatial condition $\mathcal{S}_c$ to the output image $\mathcal{I}$. During the reverse diffusion process, auxiliary shape conditions are incorporated to improve geometric consistency with the input condition. d, A residual 3D UNet with convolutional and attention blocks is implemented for the prediction of the reverse diffusion process.
• Figure 3: Image fidelity comparison of implemented methods for 3D brain MRI synthesis based on conditional cortical surfaces from the ADNI test set ($n=323$). Pix2Pix and BBDM were adapted for 3D generation. $\mathcal{R}$ and $\mathcal{E}$ indicate cortical ribbon mask and edge map inputs, respectively. a, Average vertex-wise reconstruction errors between input conditions and white matter (WM) and pial cortical surfaces reconstructed from the synthetic MRIs (lower is better). Visualizations are based on the FsAverage template. b, Structural similarity index measure (SSIM$\uparrow$) and peak signal-to-noise ratio (PSNR$\uparrow$) for whole-brain image fidelity and average symmetric surface distance (ASSD$\downarrow$, mm) for geometric consistency of right hemisphere WM and pial cortical surfaces (see Table S1 for left hemisphere values). Bar plots show the mean and standard deviation (error bar) across all samples in the test set.
• Figure 4: Brain segmentation quality and qualitative synthesis results. a, SynthSeg+-based automatic quality control (QC) scores of segmented brain structures in the synthetic MRIs. Asterisks indicate significant improvement of Cor2Vox over other methods as determined by a Wilcoxon signed rank test and Benjamini-Hochberg correction; ***: $p<0.001$, **: $p<0.01$, *: $p<0.05$. The horizontal red line indicates the recommended quality threshold for downstream processing of 0.65 billot2023synthseg+. b, Synthetic MRIs from different methods and the real counterpart, visualized together with reconstructed cortical contours. All MRIs broadly show the same cortical geometry, with differences marked by red arrows.
• Figure 5: Results from three real-world applications of Cor2Vox. a, Surface plots visualize the anatomically consistent sampling of cortices by spherical interpolation in shape space for an arbitrary pair of UKB subjects. Scatter plots show cortical thickness in the cuneus and middle temporal regions of synthetic and real data based on randomly sampled subject pairs. b, Simulation of progressive cortical atrophy; surface plots show average recovered atrophy from 124 synthetic MRIs (1 mm isotropic resolution), in dependence on the respective introduced changes. c, Augmentation of the ADNI dataset with synthetic data from Cor2Vox, based on in-house cortical shapes from frontotemporal dementia (FTD) cases. A higher multi-reference (MR)-SSIM indicates a more similar image appearance to that of ADNI. Cor2Vox accurately preserved the reduced cortical thickness in the frontotemporal region, a hallmark of FTD Du2006corticalthinningalzheimerfrontotemporal, compared to ADNI normal controls (CN).