36 Anomaly Detection with Variational Autoencoders (a) Good. (b) Shrink hole. (c) Dirty spot. (d) Discoloured line. (e) Edge erosion. (f) Print line. Figure 3.3: Examples for each defect class. Lung cancer detection, nodules from NLST 3D dataset We additionally show results for lung cancer detection in 3D images of lung nodules, as described in Santos Buitrago et al. (2018). Lung cancer detection usually requires annotated images (cancer, non-cancer) at a nodule (tumour) level, with additional information such as malignancy, diameter, spiculation, or lobulation; and a preferably large amount of samples of each class. Recent efforts (Hammack, 2017) leveraged the use of publicly available datasets with considerable nodule annotations, achieving good performance. However, this supervised approach does not seem to be easily scalable due to the lack of new, equally rich data. In this particular application, the benign nodules of the lung do not share specific characteristics. They are diverse in size, texture, shape, and location. As a consequence, differentiating between benign and malign nodules is not a trivial task even for humans. Due to the high complexity of the data, we are not
RkJQdWJsaXNoZXIy MjY0ODMw