A Case for the Score: Identifying Image Anomalies using Variational Autoencoder Gradients

TL;DR

This paper proposes using Variational Autoencoder gradients to identify image anomalies, demonstrating superior pixel-wise tumor detection performance over traditional reconstruction error methods in unsupervised settings.

Contribution

It introduces a gradient-based anomaly scoring method from VAEs, offering a theoretically grounded and more accurate alternative to reconstruction error-based detection.

Findings

VAE gradient-based ratings outperform reconstruction error in tumor detection

Achieved ROC-AUC of 0.94 on BraTS-2017 dataset

Method enhances unsupervised pixel-wise anomaly detection

Abstract

Through training on unlabeled data, anomaly detection has the potential to impact computer-aided diagnosis by outlining suspicious regions. Previous work on deep-learning-based anomaly detection has primarily focused on the reconstruction error. We argue instead, that pixel-wise anomaly ratings derived from a Variational Autoencoder based score approximation yield a theoretically better grounded and more faithful estimate. In our experiments, Variational Autoencoder gradient-based rating outperforms other approaches on unsupervised pixel-wise tumor detection on the BraTS-2017 dataset with a ROC-AUC of 0.94.