Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DISYRE: Diffusion-Inspired SYnthetic REstoration for Unsupervised Anomaly Detection

Sergio Naval Marimont, Matthew Baugh, Vasilis Siomos, Christos Tzelepis, Bernhard Kainz, Giacomo Tarroni

TL;DR

DISYRE tackles unsupervised anomaly detection in medical MRI by learning a diffusion-inspired restoration score function trained with synthetic anomalies rather than Gaussian noise. It defines a forward corruption $x_t = (1 - \\alpha_t \\cdot m)\\cdot x_0 + \\alpha_t \\cdot m \\cdot x_{fp}$ and trains a restoration network $P_{\\theta}$ to invert this process, producing an Anomaly Score (AS) from the accumulated restoration gradients. In brain MRI benchmarks, DISYRE achieves state-of-the-art-like performance on BraTS-T2 (AP ≈ 0.75) and improves ATLAS localization metrics (e.g., Dice ≈ 0.34) with multi-step inference, while BraTS-T1 remains competitive. The work demonstrates a viable diffusion-inspired approach for medical UAD that leverages synthetic degradations to generalize to real anomalies, with robust performance and practical inference strategies.

Abstract

Unsupervised Anomaly Detection (UAD) techniques aim to identify and localize anomalies without relying on annotations, only leveraging a model trained on a dataset known to be free of anomalies. Diffusion models learn to modify inputs $x$ to increase the probability of it belonging to a desired distribution, i.e., they model the score function $\nabla_x \log p(x)$. Such a score function is potentially relevant for UAD, since $\nabla_x \log p(x)$ is itself a pixel-wise anomaly score. However, diffusion models are trained to invert a corruption process based on Gaussian noise and the learned score function is unlikely to generalize to medical anomalies. This work addresses the problem of how to learn a score function relevant for UAD and proposes DISYRE: Diffusion-Inspired SYnthetic REstoration. We retain the diffusion-like pipeline but replace the Gaussian noise corruption with a gradual, synthetic anomaly corruption so the learned score function generalizes to medical, naturally occurring anomalies. We evaluate DISYRE on three common Brain MRI UAD benchmarks and substantially outperform other methods in two out of the three tasks.

DISYRE: Diffusion-Inspired SYnthetic REstoration for Unsupervised Anomaly Detection

TL;DR

DISYRE tackles unsupervised anomaly detection in medical MRI by learning a diffusion-inspired restoration score function trained with synthetic anomalies rather than Gaussian noise. It defines a forward corruption and trains a restoration network to invert this process, producing an Anomaly Score (AS) from the accumulated restoration gradients. In brain MRI benchmarks, DISYRE achieves state-of-the-art-like performance on BraTS-T2 (AP ≈ 0.75) and improves ATLAS localization metrics (e.g., Dice ≈ 0.34) with multi-step inference, while BraTS-T1 remains competitive. The work demonstrates a viable diffusion-inspired approach for medical UAD that leverages synthetic degradations to generalize to real anomalies, with robust performance and practical inference strategies.

Abstract

Unsupervised Anomaly Detection (UAD) techniques aim to identify and localize anomalies without relying on annotations, only leveraging a model trained on a dataset known to be free of anomalies. Diffusion models learn to modify inputs to increase the probability of it belonging to a desired distribution, i.e., they model the score function . Such a score function is potentially relevant for UAD, since is itself a pixel-wise anomaly score. However, diffusion models are trained to invert a corruption process based on Gaussian noise and the learned score function is unlikely to generalize to medical anomalies. This work addresses the problem of how to learn a score function relevant for UAD and proposes DISYRE: Diffusion-Inspired SYnthetic REstoration. We retain the diffusion-like pipeline but replace the Gaussian noise corruption with a gradual, synthetic anomaly corruption so the learned score function generalizes to medical, naturally occurring anomalies. We evaluate DISYRE on three common Brain MRI UAD benchmarks and substantially outperform other methods in two out of the three tasks.
Paper Structure (6 sections, 1 equation, 3 figures, 2 tables, 1 algorithm)

This paper contains 6 sections, 1 equation, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Method
  3. Experiments
  4. Discussion & Conclusion
  5. Compliance with Ethical Standards
  6. Acknowledgments

Figures (3)

  • Figure 1: 1st row: examples of anomaly generation inputs: healthy image $x_0$, shape mask $m$ and foreign patch $x_{fp}$. 2nd row: synthetic corruption forward process for different $t$ values (with $T=100$).
  • Figure 2: AP profiles as a function of $t$. Blue line shows single-step restoration. Orange line shows the cumulative AS as defined in Section \ref{['sections:method']} up-to $t$ stage ($step\_size = 25$). Single seed.
  • Figure 3: Restoration results on test images drawn from the ATLAS ATLAS, BraTS-T1, and BraTS-T2 brats, after $step\_size=25$, along with anomaly scores $AS$ in grayscale compared with the ground truth outline in red.