Cell Image Transformation Using Deep Learning

Abstract: This thesis was written at CellaVision who sells digital microscope systems, mainly used for blood analysis. Blood tests are an important part of modern health care and today digital microscopes are widely used to replace conventional microscopy. It is important that the digital images of blood cells are of high quality and that they look as they would in a traditional microscope. CellaVision has digital microscope systems with different optics, which means images from different systems do not look the same. In this thesis we investigate the possibility of transforming images between systems using neural networks. The main focus is on generative adversarial networks, also known as GANs, but we also experiment with a simple CNN and a network with a perceptual loss based on the VGG-16 network. Our results include two variations of GANs, a conditional GAN and a cyclic GAN. An advantage of the cyclic GAN is that it can be used in an unsupervised setting. It does however require a lot more memory compared to the conditional GAN. We present results from four different network setups. With these methods we have attained very good results that are better than previous tries at CellaVision. The networks are however too slow to be implemented in the actual systems today.