[Paper Review] Unsupervised Dual Adversarial Learning for Anomaly Detection in Colonoscopy Video Frames

Yuyuan Liu, Yu Tian+5 more|arXiv (Cornell University)|Oct 23, 2019

Anomaly Detection Techniques and Applications20 references4 citations

TL;DR

This paper proposes ADGAN, a novel unsupervised anomaly detection framework for colonoscopy video frames that uses a dual generative adversarial network with two generators and two discriminators to detect polyps as anomalies. Trained solely on normal frames without polyps, the model reconstructs abnormal frames poorly—producing high reconstruction errors that signal polyp presence—achieving state-of-the-art AUC of 0.7296 on a new colonoscopy dataset.

ABSTRACT

The automatic detection of frames containing polyps from a colonoscopy video sequence is an important first step for a fully automated colonoscopy analysis tool. Typically, such detection system is built using a large annotated data set of frames with and without polyps, which is expensive to be obtained. In this paper, we introduce a new system that detects frames containing polyps as anomalies from a distribution of frames from exams that do not contain any polyps. The system is trained using a one-class training set consisting of colonoscopy frames without polyps -- such training set is considerably less expensive to obtain, compared to the 2-class data set mentioned above. During inference, the system is only able to reconstruct frames without polyps, and when it tries to reconstruct a frame with polyp, it automatically removes (i.e., photoshop) it from the frame -- the difference between the input and reconstructed frames is used to detect frames with polyps. We name our proposed model as anomaly detection generative adversarial network (ADGAN), comprising a dual GAN with two generators and two discriminators. We show that our proposed approach achieves the state-of-the-art result on this data set, compared with recently proposed anomaly detection systems.

Motivation & Objective

• Address the high cost and scarcity of annotated polyp datasets by formulating polyp detection as a one-class anomaly detection problem.
• Overcome limitations of autoencoder-based methods that rely on MSE loss and fail to capture global visual features.
• Improve upon GAN-based methods that lack end-to-end training and struggle with stable image reconstruction from latent codes.
• Develop a unified, end-to-end framework that leverages adversarial training and multi-loss optimization for robust polyp detection in colonoscopy.

Proposed method

• Propose ADGAN, a dual GAN architecture with a visual generator and discriminator for image-level reconstruction, and a latent generator and discriminator for latent space mapping.
• Train the visual generator to reconstruct input images using a combined loss function comprising MSE, Wasserstein, and binary cross-entropy (BCE) losses.
• Use a two-stage training process: first pre-train the visual GAN for 80,000 iterations while freezing the latent GAN, then jointly train both GANs for 20,000 iterations.
• Generate latent codes via an encoder network that maps input images to the latent space, enabling the visual generator to reconstruct the input image.
• Compute anomaly scores using the L2 reconstruction error: $ A( extbf{x}) = ig\| extbf{x} - G_v(G_l( extbf{x})) ig\|_2^2 $, where high values indicate polyp presence.
• Optimize hyperparameters $\alpha$ and $\beta$ on the validation set within [0.1, 10] to balance loss components.

Experimental results

Research questions

• RQ1 Can a one-class learning approach trained exclusively on normal colonoscopy frames detect polyp-containing frames as anomalies with high accuracy?
• RQ2 Does combining adversarial training with a dual GAN architecture improve reconstruction fidelity and anomaly detection performance over standard autoencoders or GANs?
• RQ3 Can the proposed multi-loss function (MSE, Wasserstein, BCE) stabilize training and enhance the model’s ability to detect subtle polyp anomalies?
• RQ4 How does the end-to-end training of visual and latent GANs compare to two-stage training strategies in terms of reconstruction quality and anomaly detection AUC?
• RQ5 To what extent can the model generalize to real-world colonoscopy images with complex textures and artifacts, such as water or feces?

Key findings

• ADGAN achieves a state-of-the-art area under the ROC curve (AUC) of 0.7296 on a new colonoscopy dataset with 14,317 images, outperforming existing methods.
• The model produces significantly higher reconstruction errors for polyp-containing frames (anomaly score > 0.7296 AUC), indicating effective detection of abnormal regions.
• Visual reconstruction results show that ADGAN successfully removes polyps from input images, generating plausible healthy colon wall appearances.
• The dual GAN structure enables stable training and improved feature representation, reducing false negatives compared to standard autoencoders and GANs.
• The combination of MSE, Wasserstein, and BCE losses leads to better generalization and more discriminative anomaly scores than single-loss baselines.
• The model generalizes well to unseen test data, maintaining high performance even on complex colonoscopy frames with varying lighting and texture.