Die kardiovaskuläre Magnetresonanzbildgebung (CMR) ist eine wichtige bildgebende Methode zur Diagnose, Überwachung und Therapie von Patienten, die an Herz-Kreislauf-Erkrankungen leiden.
In der Praxis erschwert jedoch die Vielzahl der CMR-Bildgebungsparameter eine optimale Anwendung der Methodik. Entsprechend ist die Bildqualität nicht immer optimal bzw. Messungen dauern unnötig lange.
Das vorliegende Projekt hat zum Ziel, Rechenwerkzeuge und Algorithmen zu entwickeln, um die Informationskodierung, Bildrekonstruktion und Parameterbestimmung für die kardiovaskuläre Magnetresonanztomographie zu optimieren. Hierzu werden Prinzipien aus der Biophysik, Elektrotechnik, Mathematik und Informatik weiterentwickelt und synergistisch kombiniert, um letztlich die Diagnose und Behandlung von Herz-Kreislauf-Erkrankungen zu verbessern.
Im Rahmen des Projektes wird ein dreistufiger Ansatz zum Erlernen eines biophysikalisch-optimierten CMR-Designs verfolgt, bei dem 1) Simulationen und semantische Bildsynthese verwendet werden, um eine grosse Anzahl kontrollierbarer synthetischer Bilddaten zu erzeugen, 2) automatische und optimale Diagnosen unter Verwendung des Maschinellen Lernens zu ermöglichen, und 3) eine Optimierung von Datenaufnahme und -rekonstruktion zu erlernen, um diagnostische Information und Bildqualität pro Zeiteinheit zu maximieren.
Praktische Fortschritte werden auch für die experimentelle und klinische CMR-Bildgebung erwartet, indem skalierbare Trainingsdatenbanken für das Maschinelle Lernen öffentlich breitgestellt werden, die eine Quantifizierung und einen Vergleich der diagnostischen Genauigkeit über vielen Zentren hinweg erlauben.
Cardiovascular Magnetic Resonance (CMR) has become a key imaging modality to diagnose, monitor and stratify patients suffering from a wide range of cardiovascular diseases. Besides the absence of ionizing radiation and the relative operator independence, the success of CMR has been due to the multitude of image contrasts allowing to encode soft tissue properties from the micro- to the macroscale including tissue composition, tissue perfusion, metabolism, mechanical strains and blood flow among many more. At the same time, the plethora of experimental parameters involved in contrast and spatiotemporal encoding, data acquisition, image reconstruction and parameter inference makes end-to-end optimization of CMR a very challenging task. Accordingly, joint optimization of CMR pulse sequences, data sampling, image reconstruction and parameter inference has not been attempted so far. Moreover, as ground truth is missing for any in-vivo imaging situation, rigorous optimal experimental design requires provision of realistic multi-scale in-silico input data which are not yet available. In consequence, the entities of encoding, sampling, reconstruction and inference are typically “optimized” independently using experimental and/or heuristic searches and are hence not guaranteed to be optimal in terms of information yield per unit time. Information-per-unit-time optimality is, however, of critical importance in view of improving the efficiency and accuracy of imaging and for quantifying the ultimate diagnostic sensitivity and specificity.To address the challenges, a three-stage approach to learning (bio)physics-based optimal experimental CMR design is proposed using 1) Biophysical Simulation and Semantic Image Synthesis to provide large numbers and variations of paired and controllable in-silico ground truth and corresponding artificial CMR imaging data of desired complexity, 2) Parameter Inference and Disease Detection using Deep Networks based on conditional convolutional networks trained on the paired in-silico data and 3) Joint Optimization of Encoding, Sampling and Reconstruction towards optimal information-per-unit-time CMR. To this end, three work packages (WP) are defined as follows:WP 1 - Biophysical Simulation and Semantic Image Synthesis* In-vivo shape extraction and mesh generation using mode parameter learning* Multi-scale biophysical modeling of cardiac function and aortic flow to generate foreground objects* Background extraction from in-vivo data to train semantic background image synthesis* Fusion of fore- and background and input into MR physics simulation to generate realistic in-silico imagesWP 2 - Parameter Inference and Disease Detection using Deep Networks* Biophysical model parameter variation and disease modeling to generate paired in-silico datasets* Training of deep convolutional networks for parameter estimation and disease detection* Network validation using in-silico MR simulations with different contrast, resolution, artifacts and noise* Application of networks to process available pre-clinical and clinical in-vivo CMR dataWP 3 - Joint Optimization of Encoding, Sampling and Reconstruction * Formulation of encoding, sampling and reconstruction as stochastic meta-optimization problem* Definition of constraints on encoding, spatiotemporal sampling to guarantee feasibility of CMR experiments* Implementation of probabilistic relaxation to perform optimization * Integration of MR simulation and inference to perform joint optimization for optimal CMR protocol definitionThe work outlined is based on documented experience and expertise in biophysical modeling and design of CMR pulse sequences, sampling and image reconstruction schemes, image processing including the development of variational and convolutional networks for image reconstruction, image classification and parameter inference. Close links to colleagues in radiology and cardiology ensure embedding of the proposed research into a clinical context including access to CMR patient data. To this end, the research is expected to enable significant practical advances of experimental and clinical CMR by providing scalable training databases for deep network classification/segmentation applications, quantification of diagnostic accuracy relative to ground truth, CMR system identification and methods optimization delivering gains in imaging speed and information encoding efficiency towards optimal CMR.
Last updated:17.06.2022
