Many cardiovascular diseases are known to be directly related to specific blood flow phenomena. The flow field may exhibit, for example, unphysiological patterns of wall shear stress along the vessel wall triggering endothelial dysfunction and weakening of the aortic wall, while excessively high viscous shear stresses may activate thrombocytes or lead to hemolysis.
In the ascending aorta turbulent flow and high velocity jets may be found, e.g. due to anatomic particularities of the native aortic valve (bicuspid valve) and/or due to diseased aortic valves (predominantly stenotic valves). In addition, unfavourable prosthetic valve designs with flow separation may create severe turbulences downstream to the valve. Quantitative but also qualitative assessment of the flow in the ascending aorta could be an important modality to assess the risk of adverse events in the ascending aorta and the aortic arch such as dissection, aneurysms or thromboembolisms. An imaging modality could be used for risk group stratification as well as for long‐term prognosis of patients who underwent aortic valve replacement (AVR).
AVR is a common intervention to restore the proper function of the aortic valve (AV). In general, the longterm clinical outcome of AVR suffers from limited durability and biocompatibility of AV prostheses. Moreover, AVR has been related to structural deterioration of the aorta leading to aortic dissection, aneurysms and also to thromboembolism. This condition requires individual patient monitoring since undiscovered complications may be fatal or have serious consequences for quality of life (e.g. stroke). The early and precise assessment of the hemodynamics of the ascending aorta and the level of turbulence in patients contributes to an effective risk stratification and patient management with reduced risk of fatal outcomes and reduced health costs.
Presently an appropriate quantitative assessment of the flow in the ascending aorta is challenging. Existing imaging modalities including echocardiography do not offer sufficient detail and accuracy. Current echocardiographic readouts are based on a number of assumptions. For example, in clinical routine, the severity of aortic stenosis is graded using the simplified Bernoulli equation and hence the phenomenon of pressure recovery downstream of a vascular constriction is not taken into account. Consequently, the hemodynamic burden due to a stenotic valve can be overestimated with potential consequences to risk stratification and patient care. To address this limitation, time‐resolved 3D Phase‐Contrast Magnetic Resonance flow mapping (4D PC‐MRI) may be used. 4D PC‐MRI provides a vast amount of information, however, it suffers from long scan times, limited spatiotemporal resolution or intraluminal imaging artefacts. Moreover, the interpretation of time‐resolved three‐dimensional flow fields in the ascending aorta would be far too complex to be assessed and interpreted in daily clinical routine.
The proposed project “HPC‐PREDICT – High‐Performance Computing for the Prognosis of Adverse Aortic Events” aims at developing software toward a novel prognostic tool for diseases of the ascending aorta. For this purpose, modern 4D PC‐MRI of the ascending aorta shall be combined with high‐performance forward modelling of blood flow using data assimilation techniques to enhance detail and accuracy of the imaging data. The 4D PC‐MRI principle will be extended to not only encode mean velocities per voxel but also fluctuating velocities allowing to estimate the turbulent kinetic energy and the Reynolds stress tensor. In order to accommodate the added encoding dimensions within clinically feasible scan times, highly accelerated imaging concepts will be deployed in conjunction with iterative non‐linear MR image reconstruction approaches. The inherent complexity of the resulting 4D data (mean and turbulent flow field in space and time) shall be reduced through deep learning algorithms which identify landmark features in the flow field to support the prognosis of adverse events of the ascending aorta.
Because MR image reconstruction, data assimilation and deep learning involve high computational cost and clinical application requires rapid turn‐around times, high‐performance computing is an enabling technology for the proposed prognostic tool. The computational load of iterative MR image reconstruction can be addressed by multi‐threading and using parallelization. While data assimilation requires repeated realizations of the forward model (Navier‐Stokes solver) which is a massively‐parallel memory‐bandwidth3 limited task, the computational requirements for deep learning are particularly well addressed by GPUs. Therefore, hybrid supercomputing architectures, such as found in Piz Daint, are well suited for this project.
Prof. Dr. Dominik Obrist; PI; Uni Bern
Prof. Dr. Sebastian Kozerke
Prof. Dr. Ender Konukoglu
Prof. Dr. Thierry Carrel
Prof. Dr. Hendrik von Tengg-Kobligk
Prof. Dr. Rolf Krause; USI-ICS
PhD Maria Nestola; Researcher; USI-ICS
PhD Alessio Quaglino; Researcher; USI-ICS