[Paper Review] Anomaly Detection for Skin Disease Images Using Variational Autoencoder

Yuchen Lu, Peng Xu|PolyPublie (École Polytechnique de Montréal)|Jul 3, 2018

Cutaneous Melanoma Detection and ManagementMedicine13 references54 citations

TL;DR

The paper applies aVariational Autoencoder (VAE) for unsupervised anomaly detection on skin disease images, achieving notable AUCs for melanoma and AKIEC on ISIC2018 Task 3.

ABSTRACT

In this paper, we demonstrate the potential of applying Variational Autoencoder (VAE) [10] for anomaly detection in skin disease images. VAE is a class of deep generative models which is trained by maximizing the evidence lower bound of data distribution [10]. When trained on only normal data, the resulting model is able to perform efficient inference and to determine if a test image is normal or not. We perform experiments on ISIC2018 Challenge Disease Classification dataset (Task 3) and compare different methods to use VAE to detect anomaly. The model is able to detect all diseases with 0.779 AUCROC. If we focus on specific diseases, the model is able to detect melanoma with 0.864 AUCROC and detect actinic keratosis with 0.872 AUCROC, even if it only sees the images of nevus. To the best of our knowledge, this is the first applied work of deep generative models for anomaly detection in dermatology.

Motivation & Objective

• Demonstrate the potential of deep generative models, specifically VAE, for anomaly detection in dermatology.
• Train on normal skin images to detect abnormal diseases without labeled anomalies.
• Evaluate multiple VAE-based anomaly scores on ISIC2018 Task 3 dataset.

Proposed method

• Model an end-to-end Variational Autoencoder with a CNN-based encoder/decoder architecture inspired by DCGAN.
• Optimize a loss corresponding to a beta-VAE objective with a small beta to balance reconstruction and KL terms.
• Compute anomaly scores using VAE-based reconstruction likelihood and KL divergence, and compare with IWAE variants.
• Evaluate multiple anomaly scoring strategies on a dermatology dataset (ISIC2018 Task 3).
• Use 128x128 image preprocessing and ADAM optimization; latent dimension 300.

Experimental results

Research questions

• RQ1 Can a VAE trained only on normal skin images detect diverse skin diseases as anomalies?
• RQ2 Which anomaly scoring strategy (reconstruction, KL, or IWAE variants) yields best discrimination between normal and abnormal images?
• RQ3 How does VAE-based anomaly detection perform on melanoma and actinic keratosis when trained only on nevus images?
• RQ4 What are the practical limitations of using a simple prior in VAE for this task?

Key findings

• Best overall AUC achieved with reconstruction-based scores (s_vae^reconst) at 0.779 on all diseases, and up to 0.872 for AKIEC and 0.864 for MEL when evaluated individually.
• IWAE-based scores offer marginal improvements over standard VAE in this setting, with little difference observed in AUC.
• KL-based scores are less discriminative for anomaly detection under the chosen beta weighting (beta = 0.01).
• Model detects melanoma and AKIEC with high AUCs despite being trained only on nevus (normal) images.
• The study demonstrates the feasibility of applying deep generative models to dermatology anomaly detection as an unsupervised approach.