Welcome to BrainSync

Every brain scan you need. None from a patient.

Synthetic brain MRI, generated from noise.

No real study is copied, edited or anonymised. Nothing identifiable is carried forward.

Controlled lesion placement.

A segmentation mask defines where the tumour sits. The surrounding anatomy stays plausible.

Cohorts on demand.

Generate matched cases in the volume your study design requires.

Scroll
Translucent 3D rendering of a human brain

The data problem

The bottleneck is data, not architecture.

Annotated brain tumour scans are scarce, legally encumbered and slow to share. BrainSync removes the bottleneck by generating the imaging itself: full 3D MRI volumes across four sequences, with tumour placement you control. No real patient study is redistributed.

Capabilities

What the model produces

BrainSync
128³

Full 3D volumes

Complete brain volumes, not individual 2D slices.

4

Four sequences, one model

T1n, T1c, T2w and T2-FLAIR, generated together and consistent with each other.

Tumour placement by mask

Supply a segmentation mask. The lesion appears where you define it.

Fill in missing sequences

Reconstruct absent modalities from the ones you have. Any to any.

Standard research format

One NIfTI file per modality, ready for existing pipelines.

Memorisation audit

Automatically checks whether the model has memorised training patients. The central privacy question for synthetic data.

Full evaluation suite

FID, LPIPS, SSIM, precision, recall, coverage and density.

20

Sampling steps

An ODE solver produces a full volume in twenty steps.

18

Regression tests

Automated tests guard every part of the pipeline.

Architecture

How it is built

Compression
KL-VAE with PatchGAN discriminator
Generator
Diffusion Transformer · Flow Matching (OT-CFM)
Architecture
AdaLN-Zero · QK-Norm · SwiGLU · RMSNorm
Conditioning
Classifier-free guidance · dual conditioning
Solver
ODE, Euler / Heun

Method

Three stages

Generating a full 3D volume directly is intractable at clinical resolution. BrainSync works in a compressed latent space instead, which is what makes whole-volume synthesis practical.

  1. 01

    Compress

    A KL-VAE, trained against a PatchGAN discriminator, maps each 128³ multi-sequence volume into a compact latent representation.

  2. 02

    Generate

    A Diffusion Transformer trained with Flow Matching produces new latents from noise, conditioned on a tumour segmentation mask and steered with classifier-free guidance.

  3. 03

    Reconstruct

    Latents are decoded into full volumes across all four sequences and exported as one NIfTI file per modality.

Step 1: pure noise Step 2: noise beginning to form structure Step 3: a brain slice emerging Step 4: the slice nearly resolved Step 5: the finished axial brain slice
Fig. 1 From noise to structure. The generative process resolves an axial slice out of pure noise over successive steps. Illustration, not model output.

Contact

Pilot access

Enquiries from research groups and clinical AI teams are welcome. Pilot access is granted case by case.

Research preview. Model training is in progress. Early partners help shape how BrainSync is validated.

Dennis Münichsdorfner

Founder · Technology

Designed and built the BrainSync pipeline, from the latent autoencoder to the diffusion transformer, its evaluation and the memorisation audit. Your contact for all technical questions.

Tobias Berger

Co-founder · Partnerships

First point of contact for research groups and clinical teams interested in pilot access and collaboration.

Opens in your mail app — nothing is stored on this page.