Deep Learning Methods for Limited Data Problems in X-Ray Tomography

Successful medical diagnosis heavily relies on the reconstruction and analysis of images showing organs, bones, and other structures in the interior of the human body. In the last couple of years, the stored image data has increased tremendously, and also the computing power of modern GPUs experienced huge progress. Machine learning methods, and in particular deep learning methods, are on the rise to tackle advanced image reconstruction and image analysis tasks to support medical doctors in their diagnostic routines. In this chapter, we focus on the reconstruction task; especially consider tomographic imaging problems with incomplete, corrupted, or noisy data; and demonstrate how deep learning methods enable us to solve such tasks in a unified manner. We present the basic ideas of these methods assuming paired training data (supervised learning) and utilizing only feed-forward networks. In particular, we illustrate the underlying concepts for missing data problems in classical computed tomography (CT), noting that most of the concepts can be transferred to other inverse imaging problems.