Synopsis

Keywords: Flow, Brain, MRA

Quantitative time-of-flight (qTOF) magnetic resonance angiography (MRA) is a recently introduced technique that provides for simultaneous luminal and hemodynamic imaging of the intracranial arteries. We hypothesized that the application of a deep machine learning (DML) image analysis strategy to qTOF MRA data would improve agreement of intracranial arterial velocity measures with respect to phase contrast MRI. Compared to a more conventional image analysis procedure, we found that the application of DML image analysis to qTOF data improved agreement of component, total, and peak intracranial arterial flow velocity measures with respect to phase contrast MRI, and reduced calculation times by 35-fold.

Introduction

Blood flow is essential to the health and function of the brain¹. The method of time-of-flight (TOF) magnetic resonance angiography (MRA) that is in widespread clinical use to evaluate the intracranial arteries only provides for structural evaluation of the intracranial vessels and does not provide any quantitative hemodynamic information. Addressing this deficiency, the recently described approach of quantitative TOF (qTOF) MRA provides for simultaneous structural and quantitative hemodynamic evaluation of the intracranial arteries while providing similar image appearance, spatial resolution, and scan times as standard TOF MRA². However, the qTOF method to date has relied on the use of a hand-crafted computer vision procedure to quantify intracranial arterial flow velocity, which is likely suboptimal. We hypothesized that the use of a deep machine learning (DML) strategy could be used to rapidly quantify intracranial arterial flow velocity from qTOF image data. Results obtained with DML image analysis were compared to those obtained with a more conventional hand-crafted computer vision procedure.

Methods

This study was approved by our institutional review board and all participants provided written-informed consent. 15 human subjects were imaged with qTOF on a 3 Tesla MRI system (MAGNETOM Skyra^fit, Siemens Healthineers). Imaging was done using qTOF and 3D phase contrast (PC) MRI. Imaging parameters for qTOF were: TR/flip angle=21.0ms/15°, 0.58×0.58×1.0mm³ spatial resolution interpolated to 0.29×0.29×0.5mm³, scan time 4min 3s, TEs of 2.9ms (TE₁), 5.1ms (TE₂), and 7.2ms (TE₃), from which the flow-compensated TE₁ and TE₃ images were analyzed to quantify intracranial arterial flow velocity. Imaging parameters for PC were: TR/TE/flip angle=39.9ms/5.9ms/10°, 0.85×0.85×1.30mm³ spatial resolution interpolated to 0.43×0.43×0.65mm³, 60cm/s velocity encoding sensitivity, scan time 4min 4s.

Velocity quantitation with qTOF was done using conventional and DML computer vision procedures. Conventional velocity quantitation was done using hand-crafted computer vision procedure inspired from prior work², while the DML strategy was trained using spatially-registered PC velocity data as the target. The DML strategy used two 2.5D U-Nets, with one quantifying in-plane (i.e., x and y) flow velocities, and another quantifying through-plane (i.e., z) flow velocities. Measurements of component, total, and peak flow velocities were made. qTOF flow velocity measures obtained using the conventional and DML image analysis procedures were compared to PC values using intraclass correlation coefficient for absolute agreement (ICC), Pearson's correlation coefficient (r), and Bland Altman 95% limits of agreement (LOA). Computation times for both image analysis methods were measured.

Results

DML-based analysis of qTOF image data yielded spatial maps of intracranial arterial flow velocity that mimicked those obtained with PC (Figure 1). Compared to hand-crafted image analysis, DML analysis substantially improved the agreement and correlation of component flow velocities (ICC=0.966 versus 0.939, r=0.972 versus 0.940, LOA=[-5.2, +4.3] cm/s versus [-6.9, +6.5] cm/s), total flow velocities (ICC=0.835 versus 0.723, r=0.875 versus 0.744, LOA=[-6.8, +3.6] cm/s versus [-9.4, +7.1] cm/s), and peak flow velocities (ICC=0.816 versus 0.597, r=0.827 versus 0.687, LOA=[-13.5, +10.2] cm/s versus [-21.3, +10.2] cm/s). Figure 2 graphically summarizes results obtained for total flow velocity. DML analysis also reduced calculation times 35-fold (3.9±1.7 seconds versus 138±20 seconds).

Discussion

Compared to a previously published hand-crafted computer vision procedure, we found that the application of DML-based image analysis to qTOF image data improved the agreement of intracranial arterial flow velocity measures with respect to PC. The combination of rapid calculation times and improved velocimetric agreement with PC supports the application of DML to intracranial qTOF MRA. Future work will seek to refine the described DML strategy, as well as apply and validate it in patients with cerebrovascular disorders.

Conclusion

The application of deep machine learning image analysis to qTOF MRA improves the agreement of velocity measures with phase contrast MRI, and markedly reduces calculation times.

Acknowledgements

NIH award number R01EB027475