Ziele
Heute kann man sich ein alternatives Szenario vorstellen, in dem der Scanner die Erfassung automatisch für jeden Patienten anpasst, um das bestmögliche Bild zu erhalten, das den Zustand des Patienten charakterisiert, die klinischen Aufgaben unterstützt und relevante quantitative Messungen extrahiert. Im Endeffekt spielt der MRT-Scanner in diesem Szenario eine aktive Rolle bei der klinischen Untersuchung. Unser Hauptziel in diesem Projekt ist die Entwicklung von Algorithmen, mit denen ein intelligentes autonomes MR-Bildgebungssystem aufgebaut werden kann, das aktiv an klinischen Untersuchungen teilnimmt. Zu diesem Zweck werden wir hochmoderne Algorithmen für maschinelles Lernen und Signalverarbeitung entwickeln, die sich auf die Analyse von MRT-Signalen konzentrieren und das „Gehirn“ des intelligenten Bildgebungssystems darstellen. Insbesondere werden wir Algorithmen entwickeln, die MR-Bilder während der Erfassung interpretieren. Diese Algorithmen können dann Signale erzeugen, die für die Anpassung des Bildgebungsprotokolls entscheidend sind um den klinischen Zustand mit quantitativen Messungen besser zu charakterisieren.
Gesellschaftliche und wissenschaftliche Auswirkungen
Das MaRVis-Projekt wird wesentlich zum Fortschritt der Realisierung autonomer intelligenter Bildgebungssysteme beitragen, die letztendlich die Patientenversorgung verbessern werden. Um unsere Ziele zu erreichen, werden in diesem Projekt entscheidende Probleme beim maschinellen Lernen und bei der Signalverarbeitung gelöst.
Magnetic resonance imaging (MRI) is an essential tool in clinical practice for diagnosis, treatment planning and assessing therapy efficacy. Despite numerous advances in acquisition technologies, even in the most advanced centres MRI is used in a standard, non-personalised mode. An acquisition protocol is hand-picked among a set of predefined ones prior to imaging, minimally modified and executed without further adaptations.
Contrary to this mode, today's advances in machine learning and medical image computing allow us to conceive autonomous imaging systems that actively participates in the imaging by understanding the content, adapting the acquisition on the fly and searching for the best evidence to help clinical tasks. Here, we propose to develop technologies towards realising this vision.
An autonomous imaging system integrated within an MRI has the potential to revolutionise medical imaging. MRI is able to create multiple maps in 2D and 3D displaying different tissue characteristics and function at high spatial resolution and with excellent contrast between different tissue types, all without using ionising radiation. An intelligent system that can drive the acquisition autonomously can more effectively leverage this versatility to acquire better images for each patient while accelerating the overall scan, leading to better patient-care and higher throughput, even paving the way towards using MRI as a quantitative screening tool in clinical practice.While an autonomous MRI system must be composed of hardware and software components that closely interact with each other, at its core algorithms are needed to automatically interpret data as it is being acquired and extract information to guide the rest of the system.
In this project, we propose to develop the algorithms that make up this core and create a Magnetic Resonance Vision (MaRVis) system. Leveraging advances in machine learning and the release of large scale imaging datasets, we organise our work along three aims:
Aim 1. Advanced image reconstruction from partial measurements: The first aim is to develop image reconstruction algorithms for MRI that go beyond the state-of-the-art. We target algorithms that can reconstruct images with extreme undersampling ratios at a quality that allows extracting semantic information via anatomical modelling and abnormality detection, and quantify inherent uncertainties.
Aim 2. Semantic understanding and content-aware strategies for MRI:} The second aim focuses on developing algorithms that can extract semantic information during an acquisition from partially available measurements, and on designing and demonstrating strategies to leverage the extracted semantic information together with prior clinical knowledge to guide further acquisition.
Aim 3. Robust multi-contrast integrated representations: Lastly, we will develop deep learning algorithms that are robust to contrast changes and domain shifts inherent to MRI, and construct representations of the data that integrate multiple-contrasts and semantic information for better interpretation and quantification.
Successful completion of the MaRVis project will provide an algorithmic framework that opens up new avenues for how MRI can be used in clinical practice and research. By making content information available on the fly, the developed framework will enable content-aware adaptation and personalising acquisitions and motivate further developments towards acquiring optimised observations in terms of resolution, Signal-to-Noise ratio (SNR) and set of contrasts. Results in this project will take crucial steps towards realising an ultimate autonomous MRI system and can become the key motivation for future research on related hardware and cyber-physical systems, as well as other imaging systems, such as Computed Tomography.
Last updated:13.06.2022
