SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Haben Berhane
Research Staff
Ann & Robert H. Lurie Children's Hospital of Chicago
Kelly Jarvis
Postdoc
Northwestern University
Joshua Robinson, MD
Associate Professor in Pediatrics & Radiology
Ann & Robert H Lurie Children's Hospital of Chicago / Northwestern University
Cynthia Rigsby, MD
Professor of Radiology and Pediatrics
Ann & Robert H. Lurie Children's Hospital of Chicago
Michael Markl, PhD
Lester B. and Frances T. Knight Professor of Cardiac Imaging
Northwestern University
Background: 4D flow MRI provides comprehensive assessment of cardiovascular hemodynamics, including 3D visualization and quantification of time-resolved blood flow. However, extensive manual preprocessing limits its clinical application. As shown in Figure 1, 4D flow preprocessing typically involves manual selection of threshold parameters to 1) identify static tissue for estimation of phase offset errors and eddy current correction and 2) detect regions with high velocity noise for noise masking. Preprocessing can thus be time consuming and is prone to inter-observer variability. Recent developments in deep learning have shown high accuracy and robustness in image classification. The goal of this study was to develop a deep learning framework for eddy current correction (ECC) and noise masking (NM) and evaluate its performance for a fully automated 4D flow MRI preprocessing.
Methods:
A total of n=210 (100 training cases, 10 validation, and 100 testing) 4D flow MRI scans (1.5T Aera, Siemens, spatial resolution=1.2-3.2mm3, TR=37-45ms, venc=120-400cm/s) were retrospectively included. Our preprocessing workflow (Figure 1) utilizes the standard deviation over time (SD) of the 4D flow dataset to determine static tissue (SD of flow) and noise mask (SD of mag). In turn, the SD of the images were used as the input for the convolutional neural networks (CNN). Data augmentation was applied via 90o rotation and horizontal flips. As summarized in Figure 2, the networks use a Densenet architecture, in which the convolution (conv) layers were concatenated to retain as many of the features extracted by the previous layers, reducing the overall number of parameters[1]. The threshold values consisted of 8 labels, ranging 0.03-0.1. A 14-layered CNN was developed to distinguish a high [0.07-0.1] or low [0.03-0.06] threshold. Next a 13-layered CNN was used to determine the exact threshold. The weights from the first 9 layers in the initial CNN were transferred and fixed in training the latter. Separate networks were developed to determine the ECC and the NM thresholds. In total 3 CNNs were developed to determine a single threshold value: a binary classifier, a lower threshold classifier and a higher threshold classifier (Figure 2C). A dropout rate of 0.20 after every conv layer and a L2-penalty of 0.0001 were applied. Training was performed over 100 epochs with an Adam optimizer and a constant learning rate of 0.01.
Results: The ROC plot is shown in Figure 3 with an AUC of 0.97 and 0.91 for the NM and ECC CNN. The overall accuracy (softmax score threshold=0.5) was 0.93 (NM) and 0.92 (ECC) on the test cases (n=100). Computation time (GPU: Quadro P4000) per test case was 3 seconds compared to 5 minutes for manual preprocessing.
Conclusion: Deep learning can be utilized to accelerate 4D flow MRI preprocessing while maintaining excellent ECC and NM accuracy. Future direction will extend the application of deep learning to other 4D flow processing tasks such as velocity anti-aliasing or 3D segmentation.