Intracranial arterial flow velocity mapping in quantitative time-of-flight MR angiography using deep machine learning

Abstract

PurposeTo evaluate the application of deep machine learning (DML) to 3D quantitative time-of-flight (qTOF) magnetic resonance angiography (MRA) to measure blood flow velocity within the intracranial arteries. Materials and methodsIntracranial qTOF MRA was acquired in 15 subjects at 3 T. Blood flow velocity quantitation with qTOF MRA was done using a non-DML computer-vision procedure, and using convolutional DML neural networks. 3D phase contrast (PC) MRA was used as the comparator. Using PC velocity measures as the output target and qTOF two-echo source image data as inputs, DML neural networks were trained to predict component blood flow velocities. Total velocities and peak intracranial arterial blood flow velocities were computed from component velocities. ResultsCompared to non-DML image analysis, DML-based analysis of qTOF MRA image data improved agreement with PC for mean component velocity (intraclass correlation coefficient (ICC) = 0.966 versus 0.939), mean total velocity (ICC = 0.835 versus 0.723), and peak velocity (ICC = 0.816 versus 0.597), as well as narrowed the 95% Bland-Altman limits of agreement for mean component velocity ([−5.16, +4.31]cm/s versus [−6.86, +6.53]cm/s), mean total velocity ([−6.78,+3.59]cm/s versus [−9.39, +7.09]cm/s) and peak velocity ([−13.5,+10.2]cm/s versus [−21.3, +10.2]cm/s). Compared to non-DML analysis, DML image analysis reduced the root-mean-square deviation from PC velocity measures by 28%–36%, and shortened calculation times by 35-fold. ConclusionThe application of DML image analysis to intracranial qTOF MRA for velocity quantitation markedly shortened calculation times, substantially improved the agreement of component, total, and peak arterial blood flow velocities, and provided excellent agreement of hemodynamic measures with respect to 3D PC.