Figure 10.
We modify the Patch disentanglement model to 1) test the models robustness when trained with data augmented with Gaussian noise and 2) test the models response to removing the anatomical segmentation priors. While the model seems to have a small response to adding noise, removing the anatomical priors generally decreases performance. Wilcoxon signed-rank test shows that all methods are statistically significant (p-value<0.01).