A Weakly-Supervised Anomaly Detection Method via Adversarial Training for Medical Images

Abstract

Convolutional neural networks have been widely used for anomaly detection and one of their most common methods is autoencoder. The autoencoder is expected to produce lower reconstruction error for the normal data than the abnormal ones, and the reconstruction error is typically set as a measurement index for distinguishing anomalies. In practice, however, this notion is not always compatible. The autoencoder's reconstruction ability is sometimes so good that it can reconstruct anomalies with low error, resulting in the loss of anomaly detection. To address this limitation, we present a novel weakly-supervised learning method based on the generative adversarial network. The network learns the feature distribution of both normal and abnormal samples. The use of an autoencoder in the generator network allows the model to map the input image to a lower dimension vector and then remap it back to its reconstructions. The additional encoder discriminator network maps the real and generated images to their latent representations and determines whether the generated image is true or false. As a result, a higher error-index indicates that the sample is an anomaly. Experimentation on medical images from a publicly available liver dataset demonstrates the model's superiority over previous state-of-the-art approaches.