40 Anomaly Detection with Variational Autoencoders NLST 3D nodules, GAN and VAE architectures For the NLST 3D nodules dataset, we train both a GAN and a VAE. GANs are notoriously difficult to train, but after rigorous parameter tuning and training attempts we settled on the 3D WGAN-GP (Gulrajani et al., 2017) architecture shown in Figure 3.7, which was able to learn from the nodule data and produce some visually understandable results. Figure 3.7: Trained 3D WGAN-GP architecture for the generator. Seeds z are sampled from a 100-dimensional uniform distribution with coordinates between 0 and 1. We train for 100 epochs, since the loss function for the discriminator shows optimisation around epoch 50 and stops learning from epoch 60. The samples, however, keep improving visually until 100 epochs. Based on the performance of the 3D WGAN-GP architecture, we train a 3D VAE using a similar setup of 3D convolutional layers and 3D up-sampling. Figure 3.8 shows the architecture used for the encoder. Using the same 1722 normal nodules as for GANs, we train the model for 100 epochs. Figure 3.8: 3D VAE encoder architecture for NLST 3D nodules.
RkJQdWJsaXNoZXIy MjY0ODMw