Adversarial Networks: From Architecture to Practical Training

Level: Intermediate Length: 4 hours Format: In-Person Lecture Intended Audience: This intermediate-level course assumes basic knowledge in deep learning on the level of the Basic Course, SC1235 [i]Introduction to Medical Image Analysis using Convolutional Neural Networks[I]. We also assume basic programming skills in Python, as we will show code examples that participant will obtain for later review and self-learning. Description: This half-day deep dive course will guide researchers with some background knowledge, e.g. from the introductory course, SC1235 Introduction to Medical Image Analysis using Convolutional Neural Networks, through the latest literature of generative adversarial networks (GANs) and their application to medical data. First and foremost, GANs are powerful appearance models, and thus inherently bring a deep understanding of their respective domain. However, GANs can also be used to map between different domains (such as between CT and MRI) or to help training better segmentation models. Adversarial training can be introduced into several learning tasks in medical image analysis. It has been shown to help make image analysis algorithms more robust to variability in the data and to reduce the probability of failure on unseen cases. GANs in their initial implementation have been known to be hard to configure and train, but recent advances have helped them catch ground in applications of classification and segmentation, without requiring too much "witchcraft". We will introduce GANs, give an overview of their development towards the state of the art, and explain specific architectural decisions and developments that have been introduced to stabilize their training (CycleGAN, Wasserstein based loss). We will show code examples and illustrate the course content with live demonstrations on downsampled data, so that the participants gain some first-hand experience on the subject. Learning Outcomes: This course will enable you to: - explain adversarial training in general - identify several applications of GANs in medical image analysis - summarize how to implement at least one specific GAN architecture - describe typical problems in the training and how to mitigate them Instructor(s): Markus Thorsten Wenzel works on machine learning methods for medical applications since 2005 and has published more than 30 conference and journal papers on the subject. He received his PhD for his work on decision support systems for breast care. At Fraunhofer MEVIS, he is a senior scientist for cognitive medical computing. He is a funded member of the Fraunhofer Society research class "Cognitive Machines" and is experiencend in teaching and lecturing for academia and industry. He has acquired and led several international research projects. Hans Meine is a senior scientist who has been using machine learning for image analysis since 2002, and focused on various medical applications at Fraunhofer MEVIS since 2011. Since early 2016, he is organizing the internal training and coaching of Fraunhofer MEVIS staff for the new methodologies in Deep Learning, and now leads the "Image and Data Analysis" competence area that incorporates both image and non-image data. Recently, his team scored top positions in the “Liver and Tumor Segmentation” challenges at ISBI and MICCAI 2017 using Deep Learning. Volker Dicken Event: SPIE Medical Imaging 2019 Course Held: 17 February 2019