Data-driven Models and Approaches for Imaging

Abstract

Data-driven techniques are becoming popular in imaging. There is growing interest in learning signal models for various applications. We describe recent research that developed efficient, scalable, and effective data-driven models and methodologies for imaging and image processing. Efficient sparsifying transform learning methods have been proposed, often incorporating various properties such as union-of-transforms, rotation invariance, etc. Transform learning-based approaches achieve high-quality results for X-ray computed tomography (CT) or magnetic resonance image (MRI) reconstruction from limited data. We also discuss recent work on efficient methods for synthesis dictionary learning, including in combination with low-rank models. Newly proposed dictionary-based algorithms provide promising results for dynamic MRI reconstruction from limited measurements.