DELETO - Machine learning in correlative MR and high-throughput NanoCT.

Logo DELETO

Researchers: Tobias Kluth, Johannes Leuschner, Maximilian Schmidt, Pascal Fernsel
Project funding: Bundesministerium für Bildung und Forschung (BMBF), Förderschwerpunkt Mathematik für Innovationen
Project Sponsor: DESY
Partners: Martin Burger, Friedrich-Alexander-Universität Erlangen; Thomas Schuster, Universität des Saarlandes, Saarbrücken; Siemens Healthineers; ProCon X-Ray GmbH, Sarstedt; Radiologisches Institut, Universitätsklinikum Erlangen
Duration: 01.04.2020 - 31.03.2023

Machine learning (ML) and in particular large-scale neural network (NN) learning, so-called deep learning (DL), are currently among the most viral and widely discussed scientific topics, which have applications in very many research areas. In the collaborative project DELETO, the mathematical research of DL in solving inverse problems will be decisively advanced to make more accurate and efficient the reconstruction methods based on structural priors and motion correction in the field of correlative MR and high-throughput NanoCT, which are computationally expensive due to the large amount of data.

Magnetic resonance (MR) imaging is one of the most widely used medical imaging techniques. Limitations include the long measurement time and limited quantitative information associated with most MR techniques. To address these issues, this project will use learning-based techniques combined with model-driven approaches.

High-throughput nano-computed tomography (NanoCT) is based on transmission measurements of X-rays in the high-resolution range (down to 60 nm). In addition to the usual absorption imaging, phase and dark field contrasts can also be reconstructed. NanoCT allows high-precision inspection of materials such as aluminum alloys and fiber-reinforced plastics, which are of utmost economic importance (e.g. aircraft construction, automotive industry, wind turbines). Materials testing is of enormous importance to derive stress-strain relationships or to make statements about material fatigue. New requirements and materials demand ever higher resolutions with simultaneous high throughput. Sophisticated, novel mathematical methods are needed to overcome difficulties of current devices, such as inaccuracies in measurement geometry, or to develop more efficient reconstruction algorithms with Big Data measurement data at hand.

The goal is to integrate the newly developed methods into the next generation devices. To this end, we are working closely with relevant companies. For the first time, model-based and data-driven methods will be applied in the Big Data technologies of correlative magnetic resonance imaging and in high-throughput NanoCT.

The University of Bremen participates in DELETO with the subproject Invertible Residual Networks, which aims at theoretical investigations of the regulatory properties of certain network architectures, invertible residual networks (IRN).