Generative Vision Atlas

Research line · dominant

Medical: pixel space

Generate or reconstruct medical images directly in pixels, with no learned latent.

What defines membership

For inverse problems the solver needs an invertible image-to-measurement mapping, so a latent is not merely unnecessary but structurally disallowed.

How the line developed

Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.

The idea

CSGM-MRI · 2021-08core

A score-based prior for compressed-sensing MRI that generalizes out of distribution where supervised reconstruction fails.

Score-based MRI recon · 2021-10landmark

Alternates score-model sampling with k-space data consistency; the projection step is why this cannot move to a latent.

Evidence

Ktena et al. · 2024-04landmark

Pixel-space cascaded diffusion improves diagnostic accuracy and subgroup fairness across three modalities — the strongest downstream evidence in the whole medical section.

MedEdit · 2024-07strong-followup

A pixel-space denoising model that induces pathology into healthy brain MRI. Its evaluation is the notable part: a blinded neuroradiologist could not distinguish the generated strokes from real ones.

What it gets right

  • Required for inverse problems where measurements must stay consistent
  • No tokenizer means no reconstruction ceiling and no train/test latent mismatch
  • Carries the strongest downstream clinical evidence to date

Where it is weak

  • Expensive at 3D volume resolutions, which is why volumetric work left it
  • The choice is rarely argued in 2D work — it is usually inherited, not reasoned