Grounding Generative Priors for Reliable Medical Image Reconstruction

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2378 papers445 conceptsfrozen 2026-10-04

In medical imaging, there is a fundamental tension between acquisition speed and image quality. To reduce radiation dose in CT, scan time in MRI, or data rates in 3D ultrasound, researchers often "undersample" the physical signal. This leaves gaps in the data that traditional mathematical reconstructions fill with blur or streaky artifacts.

Generative models, particularly Generative Adversarial Networks (GANs) and Diffusion Probabilistic Models (DPMs), offer a tempting solution. Instead of relying solely on sparse physical measurements, these models use a "generative prior"—a deep understanding of human anatomy learned from millions of previous scans—to "hallucinate" the missing details. However, as Jelmer M. Wolterink and Salome Kazeminia have noted, this creativity is a double-edged sword: a model might synthesize a plausible-looking heart valve or a clear lung texture that simply does not exist in the patient, leading to catastrophic diagnostic errors.

The current frontier of this research is not just making images sharper, but ensuring they are "grounded"—mathematically and physically tethered to the patient's actual body.

The Power of the Generative Prior

Traditional reconstruction treats an image as a collection of pixels. Generative models treat an image as a point on a "manifold" of valid human anatomy. When we have a noisy or incomplete scan, models like those surveyed by Yuheng Fan don't just denoise the pixels; they move the image toward the nearest "plausible" version of itself.

This is exceptionally effective in ultra-sparse settings. For instance, Tristan S. W. Stevens demonstrated that diffusion models can accelerate 3D ultrasound by 10x by filling in volumetric blanks that traditional interpolation cannot handle. Similarly, in digital pathology, Zhe Xu 0012 showed that generative foundation models can restore high-frequency cellular details from out-of-focus scans by leveraging learned morphological priors.

The Hallucination Problem

The primary barrier to clinical adoption is the "hallucination"—the generation of features that satisfy the model's internal logic but violate the patient's reality. Veronika Spieker highlights this in the context of MRI motion correction: a model might "clean up" a blurred scan by inserting a standard-looking brain structure, potentially masking a real lesion or inventing a pathology.

This risk is most acute when data is extremely sparse. Matt Y. Cheung found that while diffusion priors perform "magic" with fewer than 15 X-ray projections in CT, they eventually plateau because the model begins to rely more on its learned "imagination" than the actual X-ray data. In high-stakes biological research, such as cryo-electron tomography, Brady K. Zhou notes that researchers often restrict generative outputs to visualization only, fearing that hallucinated protein structures could lead to false scientific conclusions.

Mechanisms for Anatomical Fidelity

To combat hallucinations, researchers are moving away from "black-box" image-to-image translation and toward frameworks that enforce Data Consistency (DC).

  1. Physics-Aware Iteration: Instead of a single forward pass, methods like those proposed by Matthew Tivnan for CT reconstruction integrate the generative model into a loop with the physical measurement model. The generative model suggests a detail, and the system checks if that detail is consistent with the raw X-ray data.
  2. Plug-and-Play (PnP) Priors: Yu Sun demonstrated a "Plug-and-Play" approach where a pre-trained diffusion model acts as a regularizer within a classical optimization framework. This ensures the final image is both anatomically plausible (thanks to the prior) and physically accurate (thanks to the optimization).
  3. Latent Alignment: To prevent the model from "inventing" information to fill gaps, Wencheng Han developed the Latent GSI Alignment (LGA) framework. By aligning the latent representations of standard CT scans with high-fidelity spectral data, the model learns to restore hidden details without straying from the structural grounding of the original scan.
  4. Uncertainty Guidance: Uddeshya Upadhyay introduced UP-GAN, which uses aleatoric uncertainty to identify regions where the model is "guessing." By focusing refinement on these high-uncertainty areas, the system reduces the likelihood of confident but false hallucinations.

From Passive Imaging to Active Inference

The most recent evolution in this field, described by Ruud Van Sloun, suggests a shift toward "cognitive" imaging. Here, the generative model doesn't just fix the image after the fact; it works as an active agent. By understanding what it doesn't know (uncertainty), the system can autonomously adapt its scanning strategy in real-time to collect the specific physical data needed to resolve a potential hallucination.

By shifting the role of generative models from "image creators" to "intelligent priors" that must answer to physical evidence, researchers are creating a path for AI that is both visually impressive and clinically trustworthy.

Papers in this version

Showing 60 of 378 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 605 papers
  2. Version 1 · 2026-10-04 · 378 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_anatomical_hallucination_2b0d67e6,
  title = {anatomical hallucination},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/anatomical-hallucination/rdr_2b0d67e6379f4498ab63376a782d94c2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_grounding_generative_priors_for_reliable_2b0d67e6_v1,
  title = {Grounding Generative Priors for Reliable Medical Image Reconstruction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/grounding-generative-priors-for-reliable-medical-image-reconstruction/rdr_2b0d67e6379f4498ab63376a782d94c2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}