3.6 Conclusion 49 (a) Density distribution of ELBO scores. (b) ROC curve with auROC score. Figure 3.15: VAE-based anomaly detection results for NLST 3D nodules. Likelihood values for new data can then be used to distinguish between normal data and anomalies, given a small labelled dataset of good and bad examples to determine a decision threshold. Moreover, the reconstruction capabilities of the VAE allow to visualise the difference between an image and its reconstruction, which can guide a (domain expert) observer to localise potential anomalies. Results on benchmark testing dataset MNIST show that our method indeed has the potential to separate normal data from anomalies, based on learned likelihood values. In particular, we note that the capacity of the model is important here; the model needs to be limited enough such that it only learns to represent the data it has been trained on. Too large models may generalise past the original training set and can learn to represent anomalies as well without having ever seen them. It is crucial to the anomaly framework that this does not happen. For our main use case of fault detection in surfaces of 3D-printed products, we observe that certain types of defects are easier to distinguish from normal data than others. Unsurprisingly, there appears to be a correlation between the number of pixels affected by a defect, and the performance of anomaly detection for that defect. In general, however, we show decent separation between good and anomalous data with our method, although there is also still a noticeable overlap. Moreover, we show how visualisations of the difference between images and their reconstruction can guide (human) observers to the locations of potential anomalies, which can provide a more efficient method of manually inspecting the cases that remain doubtful with our method. Lastly, we show additional results on the use case of lung cancer detection at
RkJQdWJsaXNoZXIy MjY0ODMw