[Paper Review] Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery

Thomas Schlegl, Philipp Seeböck+3 more|arXiv (Cornell University)|Mar 17, 2017

Retinal Imaging and Analysis15 references147 citations

TL;DR

The paper introduces AnoGAN, an unsupervised anomaly detection method using a DCGAN to learn normal retinal anatomy and a novel latent-space mapping with residual and discrimination losses to identify and localize anomalies in OCT images without annotated markers.

ABSTRACT

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.

Motivation & Objective

• Motivate detection of disease markers without relying on predefined marker vocabulary.
• Learn a generative model of healthy anatomical appearance from unlabeled data.
• Develop a mapping scheme from image space to latent space to identify anomalies.
• Provide an anomaly score and pixel-level anomaly localization for OCT images.
• Demonstrate detection of known and potential novel retinal markers.
• Compare to alternative unsupervised approaches and assess segmentation capability.

Proposed method

• Train a DCGAN on patches from healthy OCT images to model the normal appearance manifold.
• Map new image patches to the latent space by iteratively updating z to minimize a combined residual and discrimination loss.
• Use a residual loss to enforce similarity between x and G(z) and a discrimination loss based on discriminator features to ensure G(z) lies on the normal data manifold.
• Define an anomaly score A(x) = (1-λ)·Residual + λ·Discrimination, with λ empirically set to 0.1.
• Use feature matching-based discrimination loss to improve latent-space mapping stability.
• Evaluate on high-resolution SD-OCT data with pixel-level fluid annotations to assess detection and segmentation.

Experimental results

Research questions

• RQ1 Can a GAN-based unsupervised model learn the normal retinal anatomy to detect anomalies without marker annotations?
• RQ2 Does a learned latent-space mapping combined with residual and feature-based discrimination losses accurately identify and localize anomalies in OCT images?
• RQ3 How does AnoGAN compare to autoencoder-based or discriminator-only approaches for unsupervised anomaly detection in medical imaging?
• RQ4 Can the method detect known anomalies (e.g., retinal fluid, hyperreflective foci) and potentially novel markers?

Key findings

• The model generates realistic healthy-looking retinal patches and detects anomalies as deviations from the learned normal appearance.
• Anomaly detection using A(x) achieves high performance and both residual and discrimination components contribute to detection.
• The proposed discrimination loss based on discriminator features improves mapping to the latent space over a reference discrimination score.
• AnoGAN outperforms an adversarial autoencoder in anomaly detection performance and matches or exceeds other baselines in ROC AUC.
• The approach enables pixel-level localization of anomalies via the residual image and identifies additional lesions beyond annotated markers.
• Compared to direct discriminator-based scoring, the combined anomaly score provides robust detection across healthy and diseased images.