3.5 Results 47 see that the images look similar and are not able to create complex shapes as seen in the training data. This was also the case for less sharp images generated from simpler architectures, or with changes in the random seedz. Evenwhen the generator is able to construct simple shapes, they are very similar to each other. Figure 3.13: Four nodules generated by the 3D WGAN-GP. We use the test split, 120 normal samples and 120 anomalous samples, to calculate the loss score. We ran 100 backpropagation steps for mapping images into the latent space, and we chose λ = 0.5 in equation 3.9, after empirical experimentation. The experiment setup showed that backpropagating in the latent space was resource consuming, taking almost 30 seconds per image for 100 steps. Also, giving more weight λto one loss did not improve the resulting optimisation. Figure 3.14a shows the distribution of the results. Visually, it is
RkJQdWJsaXNoZXIy MjY0ODMw