48 Anomaly Detection with Variational Autoencoders clear that the model is not able to differentiate the distribution of normal samples from anomalous ones, as they almost entirely overlap. (a) Density distribution of anomaly scores. (b) ROC curve with auROC score. Figure 3.14: GAN-based anomaly detection results for NLST 3D nodules. Figure 3.14b shows the ROC curve and auROC score results for the GANbased method. The auROC value of 0.58 implies that the anomaly scores are not very useful for anomaly detection, and that a threshold-based classifier is barely able to perform better than random guessing. VAE results For the VAE-based anomaly detection, we used 115 normal samples and 115 anomaly samples to compute the ELBO values to be used as anomaly scores. The distribution of the values is shown in Figure 3.15a. Visually it is clear that the distributions overlap, not making a clear separation between normal and anomalous samples. Figure 3.15b shows the resulting ROC curve from thresholding on different values. We can see that even if we perform better than random guessing, the obtained scores were not good enough to make a clear distinction between normal and anomalous samples. 3.6 Conclusion We present a probabilistic anomaly detection framework that trains a Variational Autoencoder (VAE) to estimate the likelihood of normal, non-anomalous data.
RkJQdWJsaXNoZXIy MjY0ODMw