Deep Learning (DL), an area within Artificial Intelligence, has been getting a lot of attention in the area of medical imaging in the last couple of years. Several DL techniques are used for applications like disease detection, classification, and reconstruction. A problem frequently encountered is a lack of medical images from a certain disease type, which may cause the dataset to be imbalanced. In general, medical image datasets are quite small when they are compared to the natural
image datasets that are used for model comparison. The ImageNet dataset in one such dataset and it contains over 1 million training images.
A proposed solution to this data scarcity is also found in a DL technique. Generative Adversarial Networks (GANs) are able to generate new images that look realistic, but that are simulated. In the area of medical imaging, GANs have been used with several purposes, for instance to enlarge datasets, to denoise images, to turn MRI images into CT images, and for segmentation. Of interest for this assignment, are Computed Tomography (CT) images of the lungs. To increase
the performance of GANs, several researchers have experimented with image generation using semantic images. Park et al. for instance, were able to create realistic-looking images of scenery out of a semantic layout. Besides a gain in performance, it might also be easier to manipulate the content of the generated images through the semantic layout.