3D Shape-to-Image Brownian Bridge Diffusion for Brain MRI Synthesis from Cortical Surfaces

TL;DR

Cor2Vox introduces a 3D Brownian bridge diffusion model that directly translates cortical shape priors into synthetic brain MRIs by integrating detailed 3D shape representations, including cortex SDFs and auxiliary cues. The method conditions the reverse diffusion on multiple shape modalities to achieve anatomically plausible cortex geometry, validated by surface-based metrics that compare generated surfaces to ground-truth references. Across experiments and ablations, Cor2Vox delivers superior geometric accuracy and high image quality relative to 3D baselines, while preserving skull variability and enabling sub-voxel cortical atrophy simulation. This approach offers a principled, shape-aware framework for data augmentation, benchmarking, and personalized simulations in neuroimaging, with code available publicly.

Abstract

Despite recent advances in medical image generation, existing methods struggle to produce anatomically plausible 3D structures. In synthetic brain magnetic resonance images (MRIs), characteristic fissures are often missing, and reconstructed cortical surfaces appear scattered rather than densely convoluted. To address this issue, we introduce Cor2Vox, the first diffusion model-based method that translates continuous cortical shape priors to synthetic brain MRIs. To achieve this, we leverage a Brownian bridge process which allows for direct structured mapping between shape contours and medical images. Specifically, we adapt the concept of the Brownian bridge diffusion model to 3D and extend it to embrace various complementary shape representations. Our experiments demonstrate significant improvements in the geometric accuracy of reconstructed structures compared to previous voxel-based approaches. Moreover, Cor2Vox excels in image quality and diversity, yielding high variation in non-target structures like the skull. Finally, we highlight the capability of our approach to simulate cortical atrophy at the sub-voxel level. Our code is available at https://github.com/ai-med/Cor2Vox.

Figures (6)

• Figure 1: Conditioning on realistic cortical surfaces is crucial to avoid anatomical implausibilities in synthetic MRIs. (a) Generated MRIs from an unconditional 3D diffusion model. (b) Anatomical implausibilities; from top to bottom: missing central sulcus, unrealistic wide groove, scattered surface structure. (c) Real MRIs and corresponding cortical surfaces; we highlighted the characteristic central sulcus in red. (d) Our approach (Cor2Vox) incorporates real cortical surfaces to guide the generative process.
• Figure 2: Generation of cortex SDF ($\mathcal{S}_c$) and cortical ribbon mask ($\mathcal{R})$.
• Figure 3: Cor2Vox leverages a shape-to-image Brownian bridge diffusion process to learn a stochastic mapping $f_\theta$ between the shape prior $\mathcal{S}_c$ and the MRI domain $\mathcal{I}$. During the reverse diffusion process, additional shape conditions are incorporated to improve structural alignment in the generated MRIs. $f_\theta$ is modeled using a 3D UNet.
• Figure 4: Generated brain MRIs (top) and vertex-wise errors (bottom, in mm) comparing reconstructed cortical shapes to their original inputs. We show the mean error values on the cortical surface across the test set for the right hemisphere, separately for pial and white matter surfaces.
• Figure 5: We show the mean absolute difference between synthetic and original MRIs, and voxel-wise variance across synthetic MRIs across five random seeds among the whole test set, averaged along three anatomical axes, with darker colors indicating higher variability.