Real-time processing of myelin water imaging using artificial neural network
Jieun Lee1, Doohee Lee1, Joon Yul Choi1, Dongmyung Shin1, Hyeong-Geol Shin1, and Jongho Lee1

1Department of Electrical and Computer Engineering, Seoul National University, Seoul, Korea, Republic of


In this study, a real-time processing method for GRASE myelin water imaging is proposed by using an artificial neural network. Two different networks, one pairing multi-echo measurement with myelin water fraction and the other pairing multi-echo measurement with T2 distribution, were developed. Both networks took <1.5 sec for the whole brain processing (FOV = 240×180×112 mm3 and matrix = 160×120×28) with less than 5% error in white matter.


Myelin water imaging (MWI) is a sensitive imaging method for measuring myelin concentration. One of the popular sequences is a multi-echo GRASE sequence, which covers the whole cerebrum in approximately 10 min.1 The multi-echo data are processed for stimulated echo correction and regularized non-negative least-squares fitting (rNNLS) to generate the T2 distribution of myelin water and other waters from which myelin water fraction (MWF) is calculated.2-4 This data processing is computationally expensive, taking several hours of processing time. In this work, we took the advantage of the computational efficiency of artificial neural network (ANN), dramatically reducing the processing time in order to enable real-time generation of whole brain MWI.5,6 Two different networks, one pairing multi-echo data with MWF and the other pairing multi-echo data with T2 distribution, were developed (Fig. 1).



To train and evaluate the ANNs, GRASE MWI data of 35 subjects (18 healthy controls (HC) and 17 MS patients) were utilized.1 The dataset was acquired with FOV = 240×180×112 mm3, matrix size = 160×120×28, and 32 echoes (TE1 = 10 ms; echo spacing of 10 ms). As a conventional method, the rNNLS method with stimulated echo correction was applied.4 For the network training, 12 slices of 28 slices covering a large white matter volume were selected. The total number of voxels was about 70,000 per subject.

[Artificial neural network I, II]

The ANN I was designed to generate a MWF from a T2 decay curve by training the network with the 32-echo measurement as the input and the MWF from the rNNLS as the label (Fig. 1). The network consisted of 4 layers: an input layer, 4 hidden layers (96, 128, 160, and 240 neurons each), and an output layer. A leaky ReLU was used as an non-linear activation function. The loss function was defined as the mean squared error between the network output and the label data.

The ANN II generates a T2 distribution from a T2 decay curve by training the network with the 32-echo measurement as the input and the T2 distribution from the rNNLS as the label (Fig. 1). The network structure was the same as the ANN I except for the output layer (120 neurons). To enforce nonnegative values in the T2 distribution, the negative values in the output were forced to zero.

As the first step, three different subject compositions (14 HC only; 15 MS only; 7 HC and 8 MS combined) were trained to test the effects of the subject type. Then, the training data size was increased from 2 to 25 subjects to check the sufficient size for generalization. Nine subjects (4 HC and 5 MS), not included in the training, were used for this test. For the evaluation, the normalized root mean-squared error (NRMSE) was calculated for the white matter mask including lesions. Additionally, an ROI analysis for the five ROIs was performed. Lastly, the conventional MWI and the ANN II maps were compared with three different thresholds (30, 40, and 50 ms). The processing time of both methods was measured on the same quad-core computer (Intel i7-4790 CPU).


When the three subject compositions were compared, the training data with MS had smaller NRMSE (5.92%) than the data without MS (11.43%). For increasing number of training datasets, NRMSE saturated around 20 to 25 subjects. So the final networks were trained with 24 subjects (half HC and half MS). The remaining ten subjects (5 HC and 5 MS) were used for test.

The MWIs of the three methods with the difference maps in ×10 range are displayed in Fig. 2. The difference maps confirmed that the three results look similar. The average NRMSE was 4.67% in ANN I (HC:4.56%, MS:4.77%), and 4.79% in ANN II (HC:4.76%, MS:4.82%). The ROI analysis results (Table 1) show that the average MWFs in the five ROIs are not statistically different.

In Figure 3, the T2 distributions from ANN II and rNNLS are displayed, demonstrating the similarity in their shapes. Both methods revealed similar MWF maps even with the different thresholds, suggesting robustness of the ANN II in generating T2 distribution.

The processing time for the ANNs was <1.5 sec, which was shorter by 7200 times than that of the conventional method (>3 hours) on the same CPU.

Conclusion and Discussion

In this study, we developed a computationally efficient approach for estimating MWF using ANN. The result shows less than 5% of errors in <1.5 sec. Further reduction in NRMSE is expected by optimizing the network structure.


This research was supported by NRF-2017M3C7A1047864 and Brain Korea 21 Plus Project in 2018.


[1] Prasloski T, Rauscher A, MacKay AL, Hodgson M, Vavasour IM, Laule C, Mädler B. Rapid whole cerebrum myelin water imaging using a 3D GRASE sequence. Neuroimage 2012;63:533–539.

[2] Whittall KP, MacKay AL. Quantitative interpretation of magnetization transfer. Magn Reson Med 1989;84:134–152.

[3] MacKay AL, Whittall K, Adler J, Li D, Paty D, Graeb D. In vivo visualization of myelin water by magnetic resonance. Magn Reson Med 1994;31:673–677.

[4] Prasloski T, Mädler B, Xiang Q-S, MacKay AL, Jones C. Applications of stimulated echo correction to multicomponent T2 analysis. Magn Reson Med 2012;67:1803–1814.

[5] Bishop CM, Roach CM. Fast curve fitting using neural networks. Review of Scientific Instruments 1992;63(10):4450–4456.

[6] D. W. Ruck, S. K. Rogers, and M. Kabrisky. Feature selection using a multilayer perceptron. J. Neural Network Comput. 1990;1:40–48.


Figure 1. Conventional MWI using regularized NNLS (upper row), artificial neural network 1 (ANN I; middle row), and artificial neural network 2 (ANN II; bottom row). ANN I was trained to directly generate MWF whereas ANN II was trained to generate T2 distribution, from which a MWF is calculated by a threshold.

Figure 2. Comparison of the myelin water images from the three processing methods: the conventional rNNLS method (a), ANN I (b), and ANN II (c) with the difference maps displayed in ×10 range (d, and e). One slice from a healthy control (upper row) and one slice from a MS patient (lower row) are shown. The MWF maps from the three methods look similar which is confirmed by the difference maps. The average NRMSE of each method (tested by 5 HC and 5 MS) was 4.56% (HC;ANN I), 4.76% (HC;ANN II), 4.77% (MS;ANN I), and 4.82% (MS;ANN II).

Table 1. ROI analysis for the five ROIs. The results report no statistically significant difference in all ROIs.

Figure 3. Exemplary T2 distributions from rNNLS and ANN II (left) and MWF maps from three different thresholds (30, 40, and 50 ms; right). The T2 distributions from the conventional method are shown in red dashed lines and the T2 distributions from ANN II are shown in blue lines. The T2 distributions and MWF maps look similar in all cases.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)