Verbesserte Dual Energy Bildgebung mittels Maschinellem Lernen

Verbesserte Dual Energy Bildgebung mittels Maschinellem Lernen

(Third Party Funds Single)

Overall project:
Project leader:
Project members:
Start date: April 1, 2020
End date: December 31, 2020
Funding source: Europäische Union (EU)


The project aims to develop novel and innovative methods to improve visualisation and use of dual energy CT (DECT) images. Compared to conventional single energy CT (SECT) scans, DECT contains a significant amount of additional quantitative information that enables tissue characterization far beyond what is possible with SECT, including material decomposition for quantification and labelling of specific materials within tissues, creation of reconstructions at different predicted energy levels, and quantitative spectral tissue characterization for tissue analysis. However, despite the many potential advantages of DECT, applications remain limited and in specizlized clinical settings. Some reasons are that many applications are specific for the organ under investigation, require additional, manual processing or calibration, and not easily manipulated using standard interactive contrast visualisation windows available in clinical viewing stations. This is a significant disadvantage compared to conventional SECT.
In this project, we propose to develop new strategies to fuse and display the additional DECT information on a single contrast scale such that it can be visualised with the same interactive tools that radiologists are used to in their clinical routine. We will investigate non-linear manifold learning techniques like Laplacian Eigenmaps and the Sammon Mapping. Both allow extension using AI-based techniques like the newly developed user loss that allows to integrate user's opinions using forced choice experiments. This will allow a novel image contrast that will be compatible with interactive window and level functions that are rourintely used by radiologists. Furthermore, we aim at additional developments that will use deep neural networks to approximate the non-linear mapping function and to generate reconstructions that capture and display tissue specific spectral characteristics in a readily and universally useable manner for enhancing diagnostic value.