Transient wall shear stress estimation in coronary bifurcations using convolutional neural networks

Accelerated simulation methodologies for computational vascular flow modelling

“…The network with dimensionality reduction was approximately 50 times faster than the other network when performing cross-validation. Gharleghi et al [191] used a machine learning surrogate to replace a transient CFD solver in order to calculate WSS in the left main bifurcation of the coronary artery. The network requires the steady-state CFD solution for a given case as an input, but can then predict the transient WSS to an accuracy of >95normal% within 0.2 s using a CPU and 0.001 s using a GPU.…”

“…Acceleration factor 7200 a Ferdian et al [ 190 ] residual CNN super-resolution of aortic 4D flow MRI flow rate prediction accuracy . Prediction time 40–90 s Gharleghi et al [ 191 ] U-Net-style CNN transient WSS prediction in left main bifurcation of coronary arteries accuracy . Prediction time of 0.2 and 0.001 s with CPU and GPU, respectively c Li et al [ 38 ] Point-Net haemodynamics prediction before and after coronary artery bypass surgery prediction accuracy .…”

“…However, the vast amount of data required to generate accurate results could constrain these approaches, particularly in vascular flow applications where geometric data are typically derived from medical images that can be expensive to acquire and difficult to process. This is highlighted by Liang et al [ 156 ], Morales et al [ 189 ] and Gharleghi et al [ 191 ] relying upon data augmentation strategies to extend their cohorts of real patients into larger cohorts of mostly synthetic patients. While this is necessary to create sufficiently large datasets, there is a risk that the augmentation may produce unrealistic results, as demonstrated by Morales et al [ 189 ] discarding 30% of their initial training samples due to unrealistic flow features.…”

Advancements and Opportunities in Characterizing Patient-Specific Wall Shear Stress Imposed by Coronary Artery Stenting

“…Other intriguing work in this area has been carried out by Gharleghi et al, who recently applied deep learning techniques leveraging a 2 min steady simulation, artery geometry, and global features that included radii, curvature, and bifurcation angles to obtain TAWSS distributions representative of those obtained by transient CFD simulations [136]. This example study, along with related studies that have attempted to quantify [113] and reduce the time required to obtain patient-specific simulation results, represent important advancements that are critical for CFD/FSI simulations to be able to translate from the lab to clinical utility.…”

“…While the initial simulations conducted with the end goal of obtaining a realistic patient-specific simulation and its associated WSS indices may be obtained in a ~1 day, it can easily take a month or more to tune boundary conditions, to adjust material properties until deformations are consistent with available measurements, to repeatedly quantify intermediary results, and to conduct mesh and time step independence analyses in order to generate WSS indices with a high level of confidence. It is also worth noting that although their findings were exciting and focused on the left main coronary bifurcation, Gharleghi et al [136] and related work by Suk et al [137] did not include stents and applied boundary conditions that were not patient-specific. Such advancements contribute substantially to the scale and complexity of the approach, thereby likely contributing the scarcity of such studies related to stenting.…”