Source-linked AI summary
Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery
Thomas Schlegl, Philipp Seeböck, Sebastian M. Waldstein, Ursula Schmidt-Erfurth, Georg Langs
TL;DR
Existing imaging-marker detection relies on annotated examples and known-marker vocabularies, limiting discovery of relevant anomalies. AnoGAN learns normal anatomical variability from healthy data and maps new images into its latent space for anomaly scoring. On retinal OCT data, it detects known anomalies and supports segmentation, while the authors note limits from evaluating only a subset of anomaly classes.
Problem
Known-marker vocabularies and extensive annotation requirements limit imaging-based marker detection for diagnosis, disease monitoring, and treatment response.
Method
AnoGAN uses adversarial training to learn normal anatomical appearance and an improved image-to-latent mapping with generative and discriminative anomaly scoring.
Results
The approach detects known anomalies, including retinal fluid and hyperreflective foci, and identifies anomalous images and regions in retinal OCT data.
Takeaways & Limitations
Discovering anomalies at scale enables mining imaging data for candidate markers, including anomalies not represented by voxel-level annotations.
Takeaways & Limitations
Under a perfect generator and latent-space mapping, an ideal normal query has zero residual loss; quantitative evaluation is limited to a subset of anomaly classes.
Abstract
from arXiv · showhide
Obtaining models that capture imaging markers relevant for disease progression and treatment monitoring is challenging. Models are typically based on large amounts of data with annotated examples of known markers aiming at automating detection. High annotation effort and the limitation to a vocabulary of known markers limit the power of such approaches. Here, we perform unsupervised learning to identify anomalies in imaging data as candidates for markers. We propose AnoGAN, a deep convolutional generative adversarial network to learn a manifold of normal anatomical variability, accompanying a novel anomaly scoring scheme based on the mapping from image space to a latent space. Applied to new data, the model labels anomalies, and scores image patches indicating their fit into the learned distribution. Results on optical coherence tomography images of the retina demonstrate that the approach correctly identifies anomalous images, such as images containing retinal fluid or hyperreflective foci.
1 Introduction
The paper addresses limits of supervised marker detection by using unsupervised learning to model healthy anatomical appearance and identify imaging anomalies. Its GAN-based framework combines a generative model with image-to-latent mapping to detect anomalous images and regions.
- Known-marker vocabularies and extensive annotation requirements constrain the use of imaging data for diagnosis, monitoring, and treatment decisions.
- The authors propose unsupervised learning to create a rich generative model of healthy local anatomical appearance.
- The mapping and scoring approach distinguishes observations that fit the training distribution from those that do not.
- The approach uses a representative DCGAN and coupled mapping schema to discriminate normal anatomy from local anomalous appearance, making subtle anomalies at scale feasible.
- The framework combines adversarial training of a normal-appearance generative model with a coupled mapping schema for evaluating novel data.
2 Generative Adversarial Representation Learning to Identify Anomalies
The method learns normal anatomical variability with a GAN and maps query images back to latent space using residual and discriminator-feature losses. The resulting fit supports anomaly scoring and localization of anomalous regions.
- 2.1 Unsupervised Manifold Learning of Normal Anatomical Variability: A GAN learns a manifold of healthy image patches by combining a generator that produces images with a discriminator that distinguishes real from generated data.The generator maps latent noise to images, while the discriminator learns to classify real and generated examples.
- 2.2 Mapping new Images to the Latent Space: During testing, latent coordinates are iteratively optimized so the generated image most closely matches a query image while remaining on the learned normal manifold.The trained generator and discriminator remain fixed; only the latent-code coefficients are adapted.
- 2.2 Mapping new Images to the Latent Space: The mapping loss combines residual loss for visual similarity with discrimination loss for compatibility with the learned distribution of normal images.The residual term compares images directly, whereas the discrimination term uses the discriminator to enforce manifold fit.
- 2.2 Mapping new Images to the Latent Space: Feature matching replaces the discriminator’s scalar output with an intermediate feature representation when adapting latent coordinates.This uses richer discriminator information and treats the trained discriminator as a feature extractor rather than only a classifier.
- 2.3 Detection of Anomalies: The final weighted mapping loss becomes an anomaly score, with larger values indicating anomalous images and the residual image highlighting anomalous regions.Residual and discrimination scores are evaluated at the final latent-space update iteration.
3 Experiments
Experiments on retinal SD-OCT data evaluated AnoGAN's image generation, anomaly detection, and pixel-level localization using healthy training data and labeled test data. The approach generated realistic normal patches, distinguished anomalous cases, and identified retinal fluid and hyperreflective foci.
- Data, Data Selection and Preprocessing: The GAN was trained on 2D patches from 270 healthy clinical OCT volumes, with annotations withheld from training and evaluation.Positive detection labels were assigned when an image contained at least one retinal-fluid pixel.
- Evaluation: Experiments assessed realistic image generation, anomaly segmentation and discovery, and component-level detection performance against aCAE, GANR, and discriminator-output baselines.The evaluation included healthy training and test patches and diseased test images.
- Can the model generate realistic images?: The trained model generated images visually similar to normal queries but showed obvious intensity or textural differences for anomalous inputs.The qualitative assessment used healthy training/test patches and diseased test images.
- Can the model detect anomalies?: Both residual and discrimination scores separated normal from anomalous samples, while pixel-level results showed high-accuracy retinal-fluid identification and detected hyperreflective foci.The figure reports image-level score distributions and pixel-level overlays; the supplied passages do not provide the corresponding numeric AUC values.
- Can the model detect anomalies?: The proposed discrimination score outperformed the reference discrimination score, while AnoGAN and GANR achieved comparable AUCs because the residual score performed well.Good detection performance was also obtained from the discriminator output without latent-space mapping.
4 Conclusion
The paper concludes that deep generative adversarial networks can detect and segment known anomalies after unsupervised training on healthy data. This supports mining imaging data for marker candidates, although evaluation remains limited by the subset of anomaly classes assessed.
- 4 Conclusion: AnoGAN identifies retinal fluid and hyperreflective foci that were never seen during training and can segment anomalies in unseen data.The conclusion presents these findings as evidence of sensitivity and capability for anomaly segmentation.
- 4 Conclusion: Scaling anomaly discovery could support mining imaging data for marker candidates requiring future verification.The conclusion frames this as an expected application rather than a completed validation.
- 4 Conclusion: Quantitative evaluation is limited because it covers only a subset of anomaly classes and false positives do not account for novel anomalies.The stated limitation concerns interpretation of quantitative performance, not the qualitative discovery capability.