Many medical imaging tasks, like image de-noising, de-blurring, imputation, synthesis, and inference, require transformation of field data of one type to another. This transformation is often challenging to perform because it is not well defined mathematically, it is tedious perform manually, or it yields multiple likely solutions.
In this talk we present a probabilistic deep learning algorithm based on adversarial learning to solve this class of problems. We describe how, given samples from the joint distribution of two types of images (input and output), this algorithm learns the distribution of the output image conditioned on the input image and samples efficiently from this distribution. Thereafter we present applications of this algorithm in a variety of tasks including, brain extraction in MR images, image imputation in Contrast Enhanced CT of renal tumors and inferring images of mechanical properties from displacement fields acquired using ultrasound.
