Loading…
3D brain MRI synthesis from separate hemisphere and cerebellum-brainstem diffusion priors, assembled and refined into one volume with editable parts.
No providers recorded yet. Browse all providers
A synthetic brain volume can look convincing on an axial slice and still be wrong where a downstream analysis depends on it: the hemispheres drift out of correspondence, tissue interfaces blur, the cerebellum-cerebrum boundary collapses. That follows from how the volume is built. Most 3D brain generators denoise the whole head as one entangled latent, so a single network spreads its capacity across regions of very different morphology — the folded cerebral cortex against the foliated cerebellum — and the fine-scale structure of each is averaged away.
AnaDiffusion, from a seven-institution US academic collaboration led by Texas A&M University, factorizes the problem along anatomy instead. Three parts — left hemisphere, right hemisphere, and the cerebellar-brainstem complex — each get a dedicated latent diffusion prior. The generated parts are assembled in image space into a template-aligned scaffold, re-encoded by the whole-brain encoder, injected into a whole-brain latent at an intermediate noise level, and denoised onward. The whole-brain denoiser is trained specifically to resume from scaffold-injected latents, so it fills in what the parts omit — cerebrospinal fluid, background, uncertain boundaries — and harmonizes interfaces instead of overwriting the injected anatomy.
A sample is therefore not a static volume: it carries reusable part assets, so a hemisphere or the cerebellar-brainstem complex can be swapped for one from a different sample and refined back into a coherent brain. Whole-volume synthesizers such as NV-Generate-MR-Brain sample a head along one trajectory, and mask-conditioned methods need a dense segmentation map; AnaDiffusion needs neither at inference.
The implementation is PyTorch on MONAI Generative Models AutoencoderKL and diffusion UNet components, generating at 128³ for the whole brain, 64x128x128 per hemisphere and 128x96x64 for the cerebellar-brainstem complex. Autoencoders train for up to 100 epochs and are frozen before diffusion training. Sampling is a 1,000-step DDPM process with a scaled-linear beta schedule from 0.0015 to 0.0195 and 50-step DDIM at inference, with the scaffold injected at a timestep drawn uniformly from 100 to 300 in training and at 10 DDIM steps remaining at inference. Data is 1,735 T1-weighted scans from 407 ADNI subjects aged 55–93, spanning cognitively normal, mild cognitive impairment and Alzheimer's disease, skull-stripped and affinely registered to MNI152 at 1.5 mm isotropic and split by subject into 305, 41 and 61 subjects (1,288, 156 and 291 scans). SynthSeg label maps crop the part volumes offline and are never a conditioning input.
Against seven adversarial, diffusion and morphology-based baselines, AnaDiffusion records the lowest mean MedicalNet FID (reported as FID x10⁴) in all five evaluated regions: 36.16 whole brain against 40.92 for a standard latent diffusion model, 6.21 and 6.67 for the hemispheres, 1.06 for the cerebellar-brainstem complex and 0.29 in the seam band. On SynthSeg-derived morphometry it reaches the lowest cerebellar Cohen's |d| at 0.169 and second-lowest ventricular and brainstem values at 0.144 and 0.185. Disabling latent injection raises whole-brain FID to 52.27 and seam FID to 0.55.
Compositional sampling suits work needing synthetic brain MRI with control over where the variation lives: augmenting training sets for segmentation, registration or morphometry pipelines without sharing subject data; building paired cohorts that differ in one region with the rest of the anatomy held fixed; and stress-testing neuroanatomy tools against volumes whose regional composition is known by construction. For the neuroimaging groups working with restricted clinical cohorts, the part assets are the handle: an experiment is specified anatomically rather than as a hand-drawn mask.
AnaDiffusion argues that part-level structure belongs inside the generative process rather than in the conditioning signal, and shows that a scaffold injected mid-trajectory carries enough regional prior to improve local fidelity and seam coherence together. The authors' stated limitations constrain its use: the multi-stage design adds training and inference complexity over a monolithic LDM; the fixed three-part decomposition captures neither tissue classes, functional networks, nor finer anatomical hierarchies; and it assumes approximate MNI152 correspondence, so it is not validated for severe mass effect or displaced boundaries, where registration itself may fail. Evaluation is confined to ADNI T1-weighted data. The work is an arXiv preprint; its code repository is a placeholder with no release, and no weights have been published.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.