Deep Learning for Image Reconstruction: Data Requirements and Its Application to Motion-Correction in MRI

MCML Authors

Tobit Klug

Dr.

* Former Member

→ Group Reinhard Heckel
 Machine Learning and Information Processing

Abstract

In the first part of this work, we study data requirements for deep learning-based image reconstruction. We establish empirical scaling laws for supervised learning, the sample complexity of self-supervised learning and investigate test-time-training for improved data efficiency. In the second part, we propose a novel deep learning-based method as well as a novel benchmark dataset for the challenging problem of 3D MRI reconstruction under patient motion.

phdthesis Klu25

---

Dissertation

TU München. Aug. 2025

---

Authors

T. Klug

---

Links

URL

---

Research Area

 A2 | Mathematical Foundations