Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery (1703.05921v1)

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.

• The paper introduces AnoGAN, which leverages an unsupervised GAN approach to learn normal anatomical variation for detecting novel anomalies.
• It employs a novel combination of residual and discrimination losses to accurately identify retinal anomalies in medical images.
• AnoGAN outperforms baseline GAN and convolutional autoencoder methods, offering promising advances in automated marker discovery and early disease diagnosis.

Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery

The paper "Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery," authored by Thomas Schlegl, Philipp Seeböck, Sebastian M. Waldstein, Ursula Schmidt-Erfurth, and Georg Langs, presents a novel approach for the detection and quantification of anomalies in medical imaging data. The approach leverages deep convolutional generative adversarial networks (GANs) to automatically identify anomalous regions, circumventing the need for extensive manual annotation and the reliance on a predefined vocabulary of markers.

Introduction and Motivation

In the field of medical imaging, accurate detection and monitoring of disease markers are essential for diagnosis and treatment efficacy. Traditional supervised methods requiring large amounts of annotated data are limited by high annotation costs and their ability to only identify pre-known markers. This paper proposes an unsupervised method using GANs to learn a manifold of normal anatomical variability. Such an approach enables the identification of any anomaly within the data, whether previously documented or entirely novel.

Methodology

The paper introduces AnoGAN, a framework consisting of a deep convolutional GAN paired with a novel anomaly scoring scheme. The GAN is trained on healthy anatomical data to learn the complex structures and variations that occur in normal medical images. Once trained, AnoGAN can map new images to the learned latent space and identify anomalies by evaluating how well new data fits into the known distribution of healthy images.

Generative Adversarial Networks (GANs) Training:

1. Architecture:
   • The generative model captures the distribution of healthy anatomical images.
   • The discriminator model distinguishes between real medical images and those generated by the generator.
2. Loss Function:
   • A unique mapping approach was designed to transition images into the latent space.
   • Two loss components were defined:
     • Residual Loss: Measures dissimilarity between a new image and the generated image.
     • Discrimination Loss: Ensures the generated image fits the learned healthy distribution.

Results and Analysis

Qualitative and Quantitative Evaluation:

The paper used optical coherence tomography (OCT) images of the retina for evaluation. Key findings included:

1. Anomaly Detection:
   • AnoGAN was able to generate realistic images corresponding to healthy anatomical structures.
   • It accurately identified and marked retinal fluid and hyperreflective foci as anomalies with high sensitivity.
2. Quantitative Performance:
   • The algorithm achieved an AUC of 0.89 in distinguishing between normal and diseased images.
   • The results underscored that the proposed residual loss contributed significantly to accurate anomaly detection.
3. Comparison with Other Methods:
   • The paper compared AnoGAN with adversarial convolutional autoencoders (aCAE) and a baseline GAN approach (GAN_R).
   • AnoGAN demonstrated superior anomaly detection capabilities, particularly in handling high-dimensional medical images.

Implications and Future Directions

The implications of this research are multi-faceted:

1. Practical Application:
   • This unsupervised anomaly detection method can streamline the identification of disease markers in clinical environments, reducing the reliance on manual annotations.
   • The capability to discover previously undocumented anomalies could potentially unveil new markers for early disease prediction and monitoring.
2. Theoretical Impact:
   • The introduction of a coupled mapping schema adds robustness to the GAN framework, enhancing its applicability in various high-dimensional imaging tasks.
   • Future research could explore the integration of this unsupervised approach with semi-supervised learning to further enhance detection accuracy.

Conclusion

The paper provides a rigorous exploration of unsupervised anomaly detection using GANs in medical imaging, presenting a compelling case for its efficacy in marker discovery. By leveraging the strengths of generative models and novel scoring approaches, AnoGAN addresses the inherent limitations of supervised methods and opens avenues for further advancements in automated disease diagnostics. The promising results on retinal OCT images serve as a foundational step towards broader applications in medical imagery and beyond. Future research could extend this framework to other image modalities and explore its potential in real-world clinical integration.