Abstract

Inverse problems arise frequently in medical imaging applications, for instance in X-ray computed tomography (CT), where the goal is to recover the interior structural details of an underlying object from its noisy and potentially undersampled measurement. In recent years, deep learning has proved to be a transformative tool for imaging inverse problems, leading to objectively better reconstruction as compared to the classical variational framework. The talk will give a brief overview of the important deep learning-based approaches for image recovery (especially in the context of X-ray CT) while highlighting their relative merits and demerits. The key focus of the talk will be on the computational efficiency of data-driven methods, and we will discuss some of our recent contributions to developing fast learning-based approaches using ideas such as adversarial learning and unrolling stochastic optimization methods.