Preprint

Deep Bayesian REFoCUS

Recovering the multistatic data set with a generative prior

Simon Penninga, Ruud van Sloun

Eindhoven University of Technology
Department of Electrical Engineering
Correspondence: s.w.penninga@tue.nl

Resources

Watch the posterior resolve

Below is a visualization of 8 slices of the multistatic data set and the B-mode. Watch as the model transports random noise to a coherent multistatic data set across diffusion time steps.

data set x̂ · every 10th tx
loading
B-mode
loading
Interactive results

Transmit count

With a good prior, we can push subsampling to extremes. Try moving the slider to the left for a few transmit types.

Robustness to measurement noise

Here we use 32 focused beams. The model is quite good at denoising, but we find that in extreme low-SNR cases it might remove small parts of tissue.

10 dB
subject
loading

Posterior variance tracks reconstruction error

The posterior variance aligns with the loss with respect to the multistatic ground truth, showing epistemic uncertainty!

example
1 / 8
focused measurement · 8 transmits
B-mode formed directly from the eight-transmit focused measurement
posterior sample
loading posterior samples
Posterior-mean loss
Beamformed loss between the posterior mean and the multistatic ground truth
posterior variance
Beamformed posterior variance

In vivo results

In vivo multistatic datasets are hard to obtain, so the model was only trained on noiseless point-scatterer simulations. The model does not have a temporal context window, and it has never seen harmonic imaging, abberation, or even this transmit frequency! Still, hallucinations are minimal and speckle is surprisingly coherent.

transmit type
number of transmits
loading

BibTeX

@article{penninga2026refocus,
  title         = {Deep Bayesian REFoCUS},
  author        = {Penninga, Simon W. and van Sloun, Ruud J. G.},
  journal       = {arXiv preprint arXiv:2610.03419},
  year          = {2026},
  eprint        = {2610.03419},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2610.03419}
}