Pseudo-healthy Synthesis for Unsupervised Anomaly Detection

This direction is merged into another direction: reconstruction-based anomaly detection, noise-fidelity trade-off. Its last version stays readable and citable here.

Version 1 of 119 papers27 conceptsfrozen 2026-10-04

In medical imaging and industrial quality control, the most difficult anomalies to detect are those that look almost normal. A small tumor in a brain scan or a slight bend in a manufactured needle can easily be missed by standard classifiers. To solve this, researchers have converged on a powerful paradigm: Pseudo-healthy Synthesis.

The intuition is simple: if we can train a model to understand only what "healthy" or "normal" looks like, we can ask it to "fix" any image we give it. By comparing the original anomalous image with its synthesized healthy counterpart, the differences—the "residuals"—reveal the anomaly.

The Reconstruct-and-Subtract Paradigm

The core of this research direction, as summarized by Cosmin I. Bercea et al. in Denoising Diffusion Models for Anomaly Localization in Medical Images, involves a three-step process:

  1. Corruption: Take an input image (which may contain an anomaly) and partially destroy its information, typically by adding noise.
  2. Reconstruction: Use a generative model—most recently Denoising Diffusion Probabilistic Models (DDPMs)—trained exclusively on healthy data to "denoise" the image.
  3. Subtraction: Subtract the reconstructed "pseudo-healthy" image from the original. Since the model only knows how to draw healthy tissue, it will effectively "erase" the anomaly during reconstruction.

The Fundamental Noise Trade-off

The primary challenge in this field is what researchers call the "noise dilemma." As Eunwoo Kim and colleagues note in Unsupervised Anomaly Detection Using Diffusion Trend Analysis, the effectiveness of the reconstruction depends entirely on the amount of noise added.

  • Too little noise: The model retains too much information from the original image. It ends up "reconstructing" the anomaly itself, leaving nothing to find in the subtraction step.
  • Too much noise: The model loses the patient's unique anatomical identity. It might generate a perfectly healthy brain, but it won't be that patient's brain. Structural features like specific cortical folds are lost, leading to high subtraction errors even in healthy regions (false positives).

This conflict between pathology removal and identity preservation is the central theme across the literature. Yunlong Zhang et al. first framed this as a competition in their Generator Versus Segmentor (GVS) framework, replacing traditional discriminators with segmentors to better balance these two goals.

Solving the Dilemma: Masking and Adaptation

To move past the fixed-noise limitation, recent work has introduced more surgical approaches to reconstruction.

Iterative Masking: Instead of adding noise to the whole image, Ziyun Liang and J. Alison Noble developed IterMask2. This method iteratively shrinks a spatial mask toward the anomalous regions. By only reconstructing the masked areas, the model preserves the "skeleton" of the healthy anatomy while focusing its generative power on the suspected lesion.

Decoupled Sampling: Cosmin I. Bercea et al. proposed AutoDDPM, which uses a "mask, stitch, and re-sample" approach. By decoupling the noise levels used for removing the anomaly from those used for preserving healthy tissue, they allow the model to erase large lesions without blurring the rest of the scan.

Adaptive Denoising: Recognizing that different anomalies require different levels of "fixing," Hang Yao et al. introduced GLAD, which uses global and local adaptive diffusion. Similarly, Yaoxuan Feng et al. developed OmiAD, which uses adversarial distillation to calibrate the initial noise level in a single step, making the process faster and more precise.

Beyond Simple Reconstruction

The field is now moving toward more sophisticated ways of using the diffusion process. Yuan Bi and Nassir Navab demonstrated in Synomaly that we can train models to "erase" synthetic pathologies that mimic real disease morphology, improving performance in challenging modalities like ultrasound.

Finally, some researchers are moving away from reconstruction entirely. Shunsuke Sakai et al. introduced InvAD, which uses DDIM inversion to map images into a latent noise space. Instead of fixing the image, they analyze how the image "travels" through the diffusion process, achieving higher speeds and accuracy by avoiding the slow iterative reconstruction steps.

By treating anomaly detection as a problem of "translating" from a diseased state to a healthy one, these researchers have turned generative AI into a powerful diagnostic tool that requires no labeled data—only a deep understanding of what it means to be "normal."

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_pseudo_healthy_synthesis_for_unsupervise_19e4c421,
  title = {Pseudo-healthy Synthesis for Unsupervised Anomaly Detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/pseudo-healthy-synthesis-for-unsupervised-anomaly-detection/rdr_19e4c421dde54b61bdbb364ee7417764},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_pseudo_healthy_synthesis_for_unsupervise_19e4c421_v1,
  title = {Pseudo-healthy Synthesis for Unsupervised Anomaly Detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/pseudo-healthy-synthesis-for-unsupervised-anomaly-detection/rdr_19e4c421dde54b61bdbb364ee7417764/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}