Deep Generative Models in Bioimaging credit:

4 Hours

A generative model is a powerful way of learning any kind of data distribution using unsupervised learning and they have achieved tremendous success in recent years. In the context of biomedical imaging, generative models are being actively explored for many important and diverse applications that include image and video synthesis, representation learning and semi-supervised learning, domain adaptation, text to image synthesis, image compression, super-resolution, inpainting, saliency prediction, image enhancement, style transfer and texture synthesis, and image-to-image translation. Generative models also hold great potential for regularizing inverse problems that arise in biomedical image formation. This course will provide an introduction to the use of deep neural networks for generative modeling in the context of biomedical image science. An emphasis will be placed on variational autoencoders (VAEs) and generative adversarial networks (GANs), as these are currently two of the most commonly employed and efficient approaches. The use of generative models for exploring data-acquisition designs optimizing the performance of imaging systems will also be introduced. Practical details regarding the successful training and evaluation of deep generative models using biomedical image data will introduced throughout the course. This will be reinforced via homework assignments and a final project that require hands-on training of networks facilitated by allocations of GPU time. Prerequisite: BIOE 486 , BIOE 485 , BIOE 483 , BIOE 580 or consent of instructor.