Fully automated on-scanner aortic four-dimensional flow magnetic resonance imaging processing and hemodynamic analysis.

BACKGROUND

Four-dimensional (4D) flow cardiovascular magnetic resonance (CMR) is a valuable technique for evaluating cardiovascular hemodynamics, but it requires cumbersome, offline preprocessing and regional segmentation prior to quantifications or visualizations.

METHODS

The Framework for Image Reconstruction (FIRE) framework was used to integrate 4D flow processing tasks directly into the scanner reconstruction pipeline. The method builds containerized applications with standardized raw and image CMR data input/output in the open-source Magnetic Resonance Data format. In this study, deep learning models and algorithms from previous work (4D flow pre-processing, three-dimensional [3D] aorta segmentation, aorta velocity maps, quantification of aortic systolic peak velocities) were implemented in TensorFlow and executed within a containerized Python 3.6 environment on the magnetic resonance imaging (MRI) scanner, directly following the MRI data acquisition. All tasks were executed in-line using the MRI system's own computational resources. Analysis results were returned alongside the standard 4D flow CMR magnitude/phase images, available for review on-scanner console immediately after the CMR scan. In a study with 20 subjects (n = 10 patients with aortic disease, n = 10 healthy controls), FIRE performance was evaluated and compared to manual 4D flow analysis (reference standard).

RESULTS

We successfully implemented on-scanner automated 4D flow hemodynamic analysis on a 1.5T MRI system. Total on-scanner computation time for 4D flow analysis was 220 ± 35 s. Dice scores between manual vs deep learning processing (eddy current static tissue selection: 0.84 ± 0.14; noise voxel detection: 0.92 ± 0.04; aortic 3D segmentation 0.92 ± 0.06) demonstrated good to excellent pipeline performance. Bland-Altman analysis revealed a small but significant bias (0.04 m/s, p = 0.01) for peak systolic velocities between manual and deep learning processing with good limits of agreement (-0.10, 0.18 m/s) and a mean relative difference of 4% (0.8/20).

CONCLUSION

An automated 4D flow processing workflow was successfully deployed for fully automated on-scanner hemodynamic analysis with good in-line vs human performance, indicating its potential for increased workflow efficiency.