50 Anomaly Detection with Variational Autoencoders a 3D image nodule level, using both a VAE as well as a Generative Adversarial Network (GAN) to compute the anomaly score. Results show that neither of the generative models are able to capture the feature complexity of the data, resulting in weak anomaly detection performance for this application. Overall, we conclude that deep generative models can provide a suitable and sensible approach for anomaly detection, in particular in highly imbalanced settings. However, for more complex datasets, standard VAEs and GANs may fail to describe the data precise enough, leading to weaker anomaly detection performance. Further improvements to these generative models are needed, to ensure that they are general enough to accurately describe the normal data, while not being overly flexible as to include anomalous data.
RkJQdWJsaXNoZXIy MjY0ODMw