Enhancement of cerebrovascular 4D flow MRI velocity fields using machine learning and computational fluid dynamics simulation data

Abstract

Blood flow metrics obtained with four-dimensional (4D) flow phase contrast (PC) magnetic resonance imaging (MRI) can be of great value in clinical and experimental cerebrovascular analysis. However, limitations in both quantitative and qualitative analyses can result from errors inherent to PC MRI. One method that excels in creating low-error, physics-based, velocity fields is computational fluid dynamics (CFD). Augmentation of cerebral 4D flow MRI data with CFD-informed neural networks may provide a method to produce highly accurate physiological flow fields. In this preliminary study, the potential utility of such a method was demonstrated by using high resolution patient-specific CFD data to train a convolutional neural network, and then using the trained network to enhance MRI-derived velocity fields in cerebral blood vessel data sets. Through testing on simulated images, phantom data, and cerebrovascular 4D flow data from 20 patients, the trained network successfully de-noised flow images, decreased velocity error, and enhanced near-vessel-wall velocity quantification and visualization. Such image enhancement can improve experimental and clinical qualitative and quantitative cerebrovascular PC MRI analysis.

A New Definition for Intracranial Compliance to Evaluate Adult Hydrocephalus After Shunting

“…The differences between the maximum CSF velocities were less than 3.49% ( Figure 2K ). Necessarily, it does not mean our results have this error as sometimes the errors of the CINE PC-MRI method are also not negligible ( Rutkowski et al, 2021 ).…”

Medical Image-Based Computational Fluid Dynamics and Fluid-Structure Interaction Analysis in Vascular Diseases

“…Promising artificial intelligence (AI) and machine learning (ML) methods have been increasingly used in various aspects of vascular biomechanics research. These methods include imaging ( Henglin et al, 2017 ; Rutkowski et al, 2021 ); segmentation of images obtained from different imaging modalities ( Nasr-Esfahani et al, 2018 ; Guo et al, 2019 ; Livne et al, 2019 ; Zhao et al, 2019 ; Bajaj et al, 2021 ; Comelli et al, 2021 ; Tian et al, 2021 ); estimation of constitutive parameters in vivo ( Liu et al, 2019b ) or for harvested vascular tissues ( González et al, 2020 ; Liu et al, 2020 ); estimation of the zero-pressure geometry of human thoracic aorta from two pressurized geometries of the same aorta at two different blood pressure levels ( Liang et al, 2018 ); prediction of hemodynamics in human thoracic aorta trained on CFD data ( Liang et al, 2020 ) or stresses within atherosclerotic walls trained on FEA data ( Madani et al, 2019 ); computation of a probabilistic and anisotropic failure metric of the aortic wall ( Liu et al, 2021 ); and prediction of plaque vulnerability ( Cilla et al, 2012 ; Guo et al, 2021 ).…”

Non-invasive hemodynamic analysis for aortic regurgitation using computational fluid dynamics and deep learning

“…Advances in super resolution (SR) image reconstruction [11] to obtain high-resolution (HR) images from lowresolution (LR) observations are increasingly being adopted for MRI with a DL-based approach [12,13]. This approach is preferred as it not only has an advantage in spatial resolution quality over conventional SR techniques [14], but also successfully denoises flow images [15].…”

“…Other limitations include unstable and non-robust ANN architectures [16], which describe the organisational structure of the ANN's layers. The architectures plays a significant role in the performance of the DL algorithm, as well as ignoring phase/velocity aliasing error [2,15]. The aliasing error here refers to aliasing from having a velocity encoding (VENC) [19] that is too low [20] rather than other types of MRI spatial aliasing which have been explored previously and reduced [21][22][23].…”

“…The main limitation of previous work has been related to insufficient data and flow characteristics [2,15,17]. To understand the effect of additional flow characteristics in the dataset, a wider range of them were used in this study.…”