Pathology Segmentation using Distributional Differences to Images of Healthy Origin

We present a method to model pathologies in medical data, trained on data labelled on the image level as healthy or containing a visual defect. Our model not only allows us to create pixelwise semantic segmentations, it is also able to create inpaintings for the segmentations to render the pathological image healthy. Furthermore, we can draw new unseen pathology samples from this model based on the distribution in the data. We show quantitatively, that our method is able to segment pathologies with a surprising accuracy and show qualitative results of both the segmentations and inpaintings. A comparison with a supervised segmentation method indicates, that the accuracy of our proposed weakly-supervised segmentation is nevertheless quite close.