Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery

Thomas SchleglPhilipp SeeböckSebastian M. WaldsteinUrsula Schmidt-ErfurthGeorg Langs

article2017IPMI2,548 citationsBest Paper Award

Introduces AnoGAN, a deep generative adversarial framework that detects and localizes pathological markers in medical images without labeled disease data by mapping unseen scans against a learned distribution of healthy anatomy.

Listen

The article addresses the challenge of identifying disease markers in medical images such as retinal optical coherence tomography scans. Current methods rely on large annotated datasets of known markers, which restricts discovery of new indicators and demands high labeling effort. This limits the ability to exploit rich imaging data for diagnosis, progression monitoring, and treatment decisions.

The work set out to evaluate whether unsupervised learning with generative adversarial networks could model normal anatomical variability and flag anomalies as candidate markers without any labeled pathology examples.

Researchers trained a deep convolutional generative adversarial network on one million image patches extracted from 270 volumes of healthy retinas. They developed a mapping procedure from image space to the latent space that combines a residual loss and a discrimination loss based on intermediate discriminator features. They tested the resulting anomaly score on 8192 patches from 20 additional volumes, including both healthy cases and cases with retinal fluid.

The model generated realistic images of normal anatomy and produced an area under the ROC curve of 0.89 for image-level anomaly detection. It correctly identified regions of retinal fluid and also flagged hyperreflective foci never seen during training. The proposed discrimination loss improved performance over a reference approach, while the residual score alone already yielded strong results. An adversarial convolutional autoencoder performed noticeably worse.

These findings indicate that the method can surface both known and previously unannotated anomalies at scale, supporting data-driven marker discovery and reducing dependence on exhaustive manual annotation. The approach shows potential to accelerate identification of new imaging indicators for disease monitoring.

Further work should include clinical validation of newly discovered anomalies and testing on additional imaging modalities or disease types. A pilot study on larger unlabeled cohorts would help confirm generalizability before routine deployment.

The evaluation used only a modest number of test volumes and focused quantitative metrics on known fluid regions, so performance on entirely novel anomalies remains uncertain. Results are encouraging but should be interpreted with caution until larger-scale confirmation is available.

Medical University of Vienna generative adversarial networks image patches supervised machine learning unsupervised anomaly detection deep convolutional generative adversarial networks marker discovery

arXiv: 1703.05921

Paper: Generative Adversarial Networks, Ian J. Goodfellow et al. (2014). It introduces the foundational generative adversarial network (GAN) framework, which forms the core architecture that the source adapts for unsupervised anomaly detection.
Paper: Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks, Alec Radford et al. (2016). It establishes the deep convolutional GAN (DCGAN) architecture used directly by the source to model healthy anatomical variability.
Paper: Autoencoding beyond pixels using a learned similarity metric, Anders Boesen Lindbo Larsen et al. (2015). It develops the concept of using intermediate discriminator feature representations as perceptual similarity metrics, inspiring the discriminator-based loss used in the source's anomaly scoring.
Paper: Adversarial Autoencoders, Alireza Makhzani et al. (2015). It introduces adversarial autoencoders, establishing the adversarial reconstruction baseline against which the source compares its latent-space mapping GAN approach.

Paper: Deep One-Class Classification, Lukas Ruff et al. (2018). It advances deep anomaly detection by formulating an end-to-end objective (Deep SVDD) that optimizes directly for a normal-data boundary rather than relying on generative reconstruction errors.
Paper: Deep Autoencoding Gaussian Mixture Model for Unsupervised Anomaly Detection, Bo Zong et al. (2018). It extends unsupervised anomaly detection by coupling latent dimensionality reduction with joint Gaussian mixture density estimation to capture complex regular distributions.
Paper: Deep Anomaly Detection with Outlier Exposure, Dan Hendrycks et al. (2019). It improves out-of-distribution and anomaly detection by showing how exposing models to auxiliary generic outliers during training dramatically sharpens detection margins.
Paper: A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks, Kimin Lee et al. (2018). It presents a unified framework utilizing distance metrics across hidden layer representations of pre-trained networks to detect abnormal and out-of-distribution inputs.
Paper: Energy-based Out-of-distribution Detection, Weitang Liu et al. (2020). It introduces energy scores as an alternative anomaly scoring metric derived from discriminative classifiers to reliably differentiate unfamiliar test examples from normal data.

Cover for Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery

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.

Topics

generative adversarial networks image patches supervised machine learning vocabulary anomalous behavior unsupervised learning method anomaly detection methods high resolution image network architecture

Keywords

unsupervised anomaly detection deep convolutional generative adversarial networks marker discovery anomaly scoring image-level anomaly detection novelty detection adversarial training convolutional autoencoder

1 Introduction
2 Generative Adversarial Representation Learning to Identify Anomalies
2.1 Unsupervised Manifold Learning of Normal Anatomical Variability
2.2 Mapping new Images to the Latent Space
Residual Loss
Discrimination Loss
An improved discrimination loss based on feature matching
2.3 Detection of Anomalies
3 Experiments
3.1 Results
4 Conclusion
References

Knowls

Knowl 1 — AnoGAN Framework for Unsupervised Anomaly Detection
model/method
The AnoGAN framework utilizes unsupervised generative adversarial representation learning to model the distribution of normal anatomical appearance and detect anomalies in unseen medical images. A deep convolutional generative adversarial network (DCGAN) is trained exclusively on healthy image patches $\mathbf{x} \in \mathcal{X}$ , where a generator $G$ learns a mapping $G(\mathbf{z})$ from a latent space $\mathcal{Z}$ (sampled from uniform noise $\mathbf{z} \sim p_z$ ) to the image manifold of healthy anatomy, while a discriminator $D$ is trained to distinguish generated samples $G(\mathbf{z})$ from real healthy samples $\mathbf{x}$ . The network parameters are optimized via a two-player minimax game with value function:

$\min_G \max_D V(D, G) = \mathbb{E}_{\mathbf{x} \sim p_{\text{data}}(\mathbf{x})}[\log D(\mathbf{x})] + \mathbb{E}_{\mathbf{z} \sim p_z(\mathbf{z})}[\log(1 - D(G(\mathbf{z})))]$

After training on healthy data, the model parameters of $G$ and $D$ are frozen. To evaluate an unseen query image $\mathbf{x}$ , the framework finds the point $\mathbf{z}_\Gamma \in \mathcal{Z}$ whose generated image $G(\mathbf{z}_\Gamma)$ is closest to $\mathbf{x}$ in both pixel space and discriminator feature space via backpropagation in latent space. When presented with an anomalous image containing pathological alterations, the generator cannot reconstruct the anomalous features because they are absent from the learned manifold of healthy variability, resulting in high reconstruction discrepancy and high anomaly scores.
Knowl 2 — Latent Space Mapping via Feature-Matching Backpropagation
algorithm
To invert the trained generator and map an arbitrary query image $\mathbf{x}$ to its optimal latent representation $\mathbf{z}_\Gamma$ , an iterative optimization procedure is performed directly on the latent coefficients $\mathbf{z}$ using backpropagation while keeping generator $G$ and discriminator $D$ weights fixed.
Input: Query image $\mathbf{x}$ , trained generator $G$ , trained discriminator feature extractor $f(\cdot)$ , trade-off weight $\lambda$ , step count $\Gamma$ , learning rate $\eta$
Output: Optimal latent representation $\mathbf{z}_\Gamma$ , generated closest healthy image $G(\mathbf{z}_\Gamma)$
Initialize $\mathbf{z}_1 \sim \mathcal{U}(-1, 1)$
for $\gamma = 1$ to $\Gamma - 1$ do
    Generate candidate image $\hat{\mathbf{x}}_\gamma = G(\mathbf{z}_\gamma)$
    Compute residual loss $\mathcal{L}_R(\mathbf{z}_\gamma) = \sum |\mathbf{x} - \hat{\mathbf{x}}_\gamma|$
    Compute discrimination loss $\mathcal{L}_D(\mathbf{z}_\gamma) = \sum |f(\mathbf{x}) - f(\hat{\mathbf{x}}_\gamma)|$
    Compute total loss $\mathcal{L}(\mathbf{z}_\gamma) = (1 - \lambda) \cdot \mathcal{L}_R(\mathbf{z}_\gamma) + \lambda \cdot \mathcal{L}_D(\mathbf{z}_\gamma)$
    Compute gradient $g_\gamma = \nabla_{\mathbf{z}_\gamma} \mathcal{L}(\mathbf{z}_\gamma)$
    Update latent vector $\mathbf{z}_{\gamma+1} = \mathbf{z}_\gamma - \eta \cdot g_\gamma$
end for
Return $\mathbf{z}_\Gamma$ and $G(\mathbf{z}_\Gamma)$
The mapping optimization updates the latent coordinates $\mathbf{z}$ across $\Gamma = 500$ iterations using gradient descent, guiding the latent vector toward the most visually and statistically consistent healthy reconstruction of $\mathbf{x}$ .
Knowl 3 — AnoGAN Latent Mapping Loss and Anomaly Scoring Functions
equation
The latent mapping loss $\mathcal{L}(\mathbf{z}_\gamma)$ and the resulting image-level anomaly score $A(\mathbf{x})$ combine a residual term measuring pixel visual dissimilarity with a feature-matching discrimination term measuring statistical representation discrepancy.

The residual loss is defined as the $L_1$ pixel difference between the query image $\mathbf{x}$ and the generated image $G(\mathbf{z}_\gamma)$ : $\mathcal{L}_R(\mathbf{z}_\gamma) = \sum |\mathbf{x} - G(\mathbf{z}_\gamma)|$

The discrimination loss is computed from the output of an intermediate convolutional layer $f(\cdot)$ of the discriminator $D$ acting as a feature extractor, forcing generated images to match the statistical feature activations of real healthy images: $\mathcal{L}_D(\mathbf{z}_\gamma) = \sum |f(\mathbf{x}) - f(G(\mathbf{z}_\gamma))|$

The overall mapping loss at step $\gamma$ is a convex combination weighted by hyperparameter $\lambda \in [0, 1]$ : $\mathcal{L}(\mathbf{z}_\gamma) = (1 - \lambda) \cdot \mathcal{L}_R(\mathbf{z}_\gamma) + \lambda \cdot \mathcal{L}_D(\mathbf{z}_\gamma)$

Upon completing $\Gamma$ optimization steps, the image-level anomaly score $A(\mathbf{x})$ is computed using the final residual score $R(\mathbf{x}) = \mathcal{L}_R(\mathbf{z}_\Gamma)$ and discrimination score $D(\mathbf{x}) = \mathcal{L}_D(\mathbf{z}_\Gamma)$ : $A(\mathbf{x}) = (1 - \lambda) \cdot R(\mathbf{x}) + \lambda \cdot D(\mathbf{x})$

A large anomaly score $A(\mathbf{x})$ indicates that the query image $\mathbf{x}$ cannot be accurately modeled as healthy anatomy on the learned manifold.
Knowl 4 — Pixel-Level Anomaly Segmentation via Residual Images
model/method
In addition to computing an image-level anomaly score, the AnoGAN model performs localized pixel-level anomaly segmentation by computing a residual image $\mathbf{x}_R$ :

$\mathbf{x}_R = |\mathbf{x} - G(\mathbf{z}_\Gamma)|$

where $\mathbf{x}$ is the input query image and $G(\mathbf{z}_\Gamma)$ is the image reconstructed from the optimal latent vector $\mathbf{z}_\Gamma$ found after $\Gamma$ backpropagation iterations. Because the generator $G$ has only been trained on healthy anatomical structures, it can only synthesize healthy representations that reside on the learned normal manifold $\mathcal{X}$ . In regions where the input $\mathbf{x}$ exhibits non-healthy structures (such as retinal fluid cysts or hyperreflective foci), $G(\mathbf{z}_\Gamma)$ reconstructs the corresponding healthy tissue morphology. The absolute difference across pixels in $\mathbf{x}_R$ isolates and highlights anomalous lesions without requiring prior pixel-level annotations during training.
Knowl 5 — Retinal SD-OCT Dataset and Preprocessing Protocol
experimental setup
The AnoGAN framework was evaluated on clinical high-resolution spectral-domain optical coherence tomography (SD-OCT) volumes of the retina. Each volume contains 49 B-scans (zx-plane image slices) with dimensions $496 \times 512 \times 49$ voxels in the $z$ -, $x$ -, and $y$ -directions, respectively.

Preprocessing steps:
1. Gray-value intensities were normalized to the range $[-1, 1]$ .
2. The OCT volumes were resized in the $x$ -direction to a spatial scale of $22\,\mu\text{m}$ , yielding approximately 256 columns per B-scan.
3. The retinal slab was segmented between top and bottom intraretinal boundaries using an automated layer segmentation algorithm and flattened to remove variations in anatomical curvature, orientation, and thickness.
4. From the preprocessed healthy volumes of 270 subjects (selected based on the absence of fluid), 1,000,000 2D patches of size $64 \times 64$ pixels were randomly extracted for unsupervised GAN training.
5. The test set comprised 8,192 image patches ( $64 \times 64$ pixels) extracted from 10 unseen healthy cases and 10 unseen pathological cases exhibiting intraretinal fluid. Voxel-wise annotations of fluid from clinical retina experts served as ground truth for statistical evaluation, where a patch was labeled anomalous if it contained at least one pixel of annotated fluid.
Knowl 6 — AnoGAN DCGAN Architecture and Optimization Configuration
experimental setup
The GAN used in AnoGAN adapts the Deep Convolutional GAN (DCGAN) architecture for single-channel grayscale $64 \times 64$ pixel inputs:
- Generator $G$ : Composed of 4 fractionally-strided convolution layers with $5 \times 5$ filters and feature map depths of 512, 256, 128, and 64, followed by an output layer producing a $64 \times 64$ grayscale patch.
- Discriminator $D$ : Composed of 4 standard convolution layers with $5 \times 5$ filters and matching intermediate channel depths of 64, 128, 256, and 512, mapped to a scalar prediction. The intermediate feature map $f(\cdot)$ for feature-matching loss is extracted from the last convolutional layer before the final scalar output.
- Training Optimization: Trained for 20 epochs using the Adam stochastic optimizer.
- Latent Mapping Optimization: For testing query images, $\Gamma = 500$ backpropagation steps were run per image patch.
- Loss Weighting: The trade-off parameter between residual and discriminator feature losses was set to $\lambda = 0.1$ for all mapping and scoring computations.

Knowl 7 — Performance Comparison Across Anomaly Detection Models

data/table

The quantitative detection performance of AnoGAN was evaluated on the SD-OCT test set against an adversarial convolutional autoencoder (aCAE), direct discriminator classification output ( $P_D$ ), and a reference GAN mapping baseline ( $\text{GAN}_R$ ) that uses cross-entropy discriminator classification loss $\mathcal{L}_{\hat{D}}(\mathbf{z}_\gamma) = \sigma(D(G(\mathbf{z}_\gamma)), 1)$ instead of feature matching.

Method	Precision	Recall	Sensitivity	Specificity	AUC
aCAE	0.7005	0.7009	0.7011	0.6659	0.73
$P_D$	0.8471	0.5119	0.5124	0.8970	0.72
$\text{GAN}_R$	0.8482	0.7631	0.7634	0.8477	0.88
AnoGAN	0.8834	0.7277	0.7279	0.8928	0.89

Metrics were computed at the optimal cut-off point on the receiver operating characteristic (ROC) curves identified using Youden's index. AnoGAN achieved the highest AUC (0.89) and precision (0.8834). While the autoencoder-based model (aCAE) provided direct latent mappings, it demonstrated lower performance (AUC 0.73) due to over-adapting to anomalous images. The feature-matching discrimination score in AnoGAN ( $D(\mathbf{x})$ , AUC 0.74) also substantially outperformed the scalar cross-entropy discrimination score ( $\hat{D}(\mathbf{x})$ , AUC 0.50), demonstrating the advantage of intermediate discriminator feature matching.

Knowl 8 — Unsupervised Discovery of Unannotated Retinal Pathologies
empirical result
When evaluated on clinical SD-OCT test cases, AnoGAN accurately detected both targeted fluid anomalies and previously unannotated disease markers. While training was conducted exclusively on fluid-free healthy scans and quantitative evaluations used expert annotations of retinal fluid, the residual image $x_R$ and anomaly score $A(\mathbf{x})$ correctly localized and flagged other distinct retinal lesions, specifically hyperreflective foci (HRF). Because HRF represent pathological structural deviations absent from the learned manifold of healthy retinal anatomy, the generator produced healthy tissue in their place, creating localized residual peaks. This demonstrates that unsupervised manifold learning with GANs can discover novel or unannotated biomarker candidates without task-specific supervisory labels.

Coverage note — None was omitted; all contributed methodology, algorithm steps, loss formulations, architectural details, dataset configurations, and quantitative/qualitative results have been captured.

References

1.Del Giorno, A., Bagnell, J.A., Hebert, M.: A discriminative framework for anomaly detection in large videos. In: ECCV, Springer (2016) 334–349
2.Matteoli, S., Diani, M., Theiler, J.: An overview of background modeling for detection of targets and anomalies in hyperspectral remotely sensed imagery. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 7(6) (2014) 2317–2336
3.Carrera, D., Boracchi, G., Foi, A., Wohlberg, B.: Detecting anomalous structures by convolutional sparse models. In: 2015 International Joint Conference on Neural Networks (IJCNN), IEEE (2015) 1–8
4.Erfani, S.M., Rajasegarar, S., Karunasekera, S., Leckie, C.: High-dimensional and large-scale anomaly detection using a linear one-class SVM with deep learning. Pattern Recognition 58 (2016) 121–134
5.Pimentel, M.A., Clifton, D.A., Clifton, L., Tarassenko, L.: A review of novelty detection. Signal Processing 99 (2014) 215–249
6.Venhuizen, F.G., van Ginneken, B., Bloemen, B., van Grinsven, M.J., Philipsen, R., Hoyng, C., Theelen, T., S´anchez, C.I.: Automated age-related macular degen- eration classification in oct using unsupervised feature learning. In: SPIE Medical Imaging, International Society for Optics and Photonics (2015) 94141I–94141I
7.Schlegl, T., Waldstein, S.M., Vogl, W.D., Schmidt-Erfurth, U., Langs, G.: Predict- ing semantic descriptions from medical images with convolutional neural networks. In: International Conference on Information Processing in Medical Imaging. Vol- ume 24., Springer (2015) 437–448
8.Seeb¨ock, P., Waldstein, S., Klimscha, S., Gerendas, B.S., Donner, R., Schlegl, T., Schmidt-Erfurth, U., Langs, G.: Identifying and categorizing anomalies in retinal imaging data. NIPS 2016 MLHC workshop. preprint arXiv:1612.00686 (2016)
9.Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in Neural Information Processing Systems. (2014) 2672–2680
10.Denton, E.L., Chintala, S., Fergus, R., et al.: Deep generative image models using a laplacian pyramid of adversarial networks. In: Advances in neural information processing systems. (2015) 1486–1494
11.Donahue, J., Kr¨ahenb¨uhl, P., Darrell, T.: Adversarial feature learning. arXiv:1605.09782 (2016)
12.Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv:1511.06434 (2015)
13.Yeh, R., Chen, C., Lim, T.Y., Hasegawa-Johnson, M., Do, M.N.: Semantic image inpainting with perceptual and contextual losses. arXiv:1607.07539 (2016)
14.Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training GANs. In: Advances in Neural Information Pro- cessing Systems. (2016) 2226–2234
15.Garvin, M.K., Abr`amoff, M.D., Wu, X., Russell, S.R., Burns, T.L., Sonka, M.: Automated 3-D intraretinal layer segmentation of macular spectral-domain optical coherence tomography images. Transactions on Medical Imaging, IEEE 28(9) (2009) 1436–1447
16.Pathak, D., Kr¨ahenb¨uhl, P., Donahue, J., Darrell, T., Efros, A.A.: Context en- coders: Feature learning by inpainting. CoRR abs/1604.07379 (2016)
17.Kingma, D., Ba, J.: Adam: A method for stochastic optimization. arXiv:1412.6980 (2014)
18.Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G.S., Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., Levenberg, J., Man´e, D., Monga, R., Moore, S., Murray, D., Olah, C., Schuster, M., Shlens, J., Steiner, B., Sutskever, I., Talwar, K., Tucker, P., Vanhoucke, V., Vasudevan, V., Vi´egas, F., Vinyals, O., Warden, P., Wattenberg, M., Wicke, M., Yu, Y., Zheng, X.: TensorFlow: Large-scale machine learning on heterogeneous systems (2015) Software available from tensorflow.org.

Citation

MLA

Schlegl, T., et al. “Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery”. Lecture Notes in Computer Science, Springer International Publishing, 2017, pp. 146–57, https://doi.org/10.1007/978-3-319-59050-9_12.

APA

Schlegl, T., Seeböck, P., Waldstein, S. M., Schmidt-Erfurth, U., & Langs, G. (2017). Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery. In Lecture Notes in Computer Science (pp. 146–157). Springer International Publishing. https://doi.org/10.1007/978-3-319-59050-9_12

Chicago

Schlegl, T., P. Seeböck, S. M. Waldstein, U. Schmidt-Erfurth, and G. Langs. 2017. “Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery”. In Lecture Notes in Computer Science. Springer International Publishing. https://doi.org/10.1007/978-3-319-59050-9_12.

Harvard

Schlegl, T. et al. (2017) “Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery”, Lecture Notes in Computer Science. Springer International Publishing, pp. 146–157. Available at: https://doi.org/10.1007/978-3-319-59050-9_12.

Vancouver

1. Schlegl T, Seeböck P, Waldstein SM, Schmidt-Erfurth U, Langs G (2017) Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery. In: Lecture Notes in Computer Science. Springer International Publishing, pp 146–157

BibTeX

@inbook{Schlegl_2017, title={Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery}, ISBN={9783319590509}, ISSN={1611-3349}, url={http://dx.doi.org/10.1007/978-3-319-59050-9_12}, DOI={10.1007/978-3-319-59050-9_12}, booktitle={Information Processing in Medical Imaging}, publisher={Springer International Publishing}, author={Schlegl, Thomas and Seeböck, Philipp and Waldstein, Sebastian M. and Schmidt-Erfurth, Ursula and Langs, Georg}, year={2017}, pages={146–157} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — AnoGAN Framework for Unsupervised Anomaly Detection

Knowl 2 — Latent Space Mapping via Feature-Matching Backpropagation

Knowl 3 — AnoGAN Latent Mapping Loss and Anomaly Scoring Functions

Knowl 4 — Pixel-Level Anomaly Segmentation via Residual Images

Knowl 5 — Retinal SD-OCT Dataset and Preprocessing Protocol

Knowl 6 — AnoGAN DCGAN Architecture and Optimization Configuration

Knowl 7 — Performance Comparison Across Anomaly Detection Models

Knowl 8 — Unsupervised Discovery of Unannotated Retinal Pathologies

References

Citation

BibTeX

Access the Paper