3982

Neural network for fast quantitative magnetization transfer imaging
Donghyun Kim1, Jae-Woong Kim1, Seung Hong Choi2, and Sung-Hong Park1

1Korea Advanced Institute of Science and Technology, Daejeon, Korea, Republic of, 2Seoul National University Hospital, Seoul, Korea, Republic of

Synopsis

Quantitative magnetization transfer (qMT) imaging provides quantitative information of macromolecular properties of tissues, but requires a long scan time. In this study, a neural network is proposed for the acceleration of qMT imaging. The network was trained to output the full 12 MT images (6 offset frequencies, 2 RF powers) from an input of only 4 MT images (2 offset frequencies, 2 RF powers). The qMT imaging with the neural network showed results comparable to those from the conventional qMT imaging with the full 12 MT images, indicating reduction in scan time by a factor of 3.

INTRODUCTION

The quantitative magnetization transfer (qMT) imaging provides quantitative information of the magnetization exchange between the macromolecular pool and the free water pool [1]. However, the long scan time is one of the obstacles for qMT imaging to be routinely used in clinics. Recently, deep learning approaches are suggested to solve many issues of biomedical imaging [2,3]. In this study, a neural network is proposed for the acceleration of qMT imaging by taking 4 MT images as input and predicting the full 12 MT images as output. This will reduce the acquisition time by a factor of 3 and hence increase chances of clinical translation and decrease undesired artifacts such as motion artifacts. The network was tested on conventional qMT and inter-slice qMT datasets [4].

METHODS

All datasets were obtained from 7 healthy volunteers using a Siemens 3T Tim Trio system (Siemens Medical Solutions, Erlangen, Germany). Balanced SSFP was used as readout for both conventional and inter-slice qMT with scan parameters of TE=2.275ms, TR=4.55ms, matrix size=128´128, slice thickness=5mm, FOV=220´220 mm², and number of slices=1 (conventional) and 25 (inter-slice). T1 and T2 maps were obtained by inversion recovery bSSFP and multi-echo spin echo, respectively. For both qMT imaging methods, total 12 MT images were acquired at 6 off-resonance frequencies (2, 3, 5, 9, 15, 25kHz) with two MT saturation RF flip angles (30°, 75°). Since MT effects at 25 kHz off-resonance were negligible, the images of 25 kHz off-resonance were considered as MT free images, by which all the data was normalized.

The network was trained to take 4 MT images as input and to predict the 8 MT images at the remaining off-resonance frequencies as output. The network was composed of 6 fully-connected layers with the sigmoid activation function. A dense connection [5] was applied between layer1 and layer3 (Fig. 1).

For training, Adam optimizer [6] with learning rate 0.001 was used to reduce the mean squared error (MSE) with scaling factor 0.01. Seven-fold cross-validation was used to validate the results. Early stopping was applied after 100 epochs for conventional qMT and 2000 epochs for inter-slice qMT. To train the conventional qMT data, the network was pre-trained with the inter-slice data to compensate for the small data size.

The outputs from the various input choices were compared to choose the best frequencies for the 4 input. The two MT images from 25 kHz offset were always included for the data normalization and the remaining one off-resonance frequency for each FA was searched. The qMT fitting was conducted to find the exchange rate (kf) and the pool fraction (F) for the data from the deep learning networks and also using only 4, 6, 8 and 10 MT images in the conventional way with no deep learning, through the database approach described in Kim. J-W et al. [4]. The qMT fitting applied to the original 12 MT images was used as ground truth. The performance was evaluated by peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and normalized root mean squared error (NRMSE).

RESULTS and DISCUSSION

For the comparison of input frequency choice, the set of 2 kHz and 25 kHz showed the best results for both flip angles (Fig. 2), so this frequency set was used in the remaining experiments.

Figures 3 and 4 are the predicted MT images from a representative subject. PSNR / SSIM / NRMSE for the conventional and inter-slice qMT datasets were $$$45.2\pm1.8$$$ / $$$0.9936\pm0.0006$$$ / $$$0.0056\pm0.0011$$$ and $$$39.4\pm1.3$$$ / $$$0.9801\pm0.0029$$$ / $$$0.0110\pm0.0017$$$, respectively. Network for conventional qMT showed better performance than that for the inter-slice qMT, presumably due to higher SNR of the images. Figures 4 and 5 show the qMT fitting results. For the conventional qMT, the network trained with the 4 input MT images showed performance comparable to those of the 8-MT images processed in the conventional way with no deep learning. On the other hand, for the inter-slice qMT the network trained with the 4 input MT images showed higher performance than all the conventional approach including the case of the 10 MT images. The difference in performance of the deep learning networks between the conventional and inter-slice qMT maps might be related to data size used for training.

CONCLUSION

The proposed deep learning networks could reduce the scan time by a factor of 3 with relatively high SSIM and PSNR and low NRMSE, indicating qMT results of the networks close to those of the ground truth. Further studies are necessary for the network to be improved by training with larger data from various subjects including patients.

Acknowledgements

No acknowledgement found.

References

1. HENKELMAN, R. Mark, et al. Quantitative interpretation of magnetization transfer. Magnetic resonance in medicine, 1993, 29.6: 759-766. 2. RONNEBERGER, Olaf; FISCHER, Philipp; BROX, Thomas. U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015. p. 234-241. 3. KIM, Ki Hwan; DO, Won‐Joon; PARK, Sung‐Hong. Improving resolution of MR images with an adversarial network incorporating images with different contrast. Medical physics, 2018. 4. Kim. J-W, et al. Rapid whole brain qMT imaging with inter-slice MT effects and database-driven fitting approach. 26th Annual Meeting, ISMRM 2018; Program number: 2754 5. HUANG, Gao, et al. Densely Connected Convolutional Networks. In: CVPR. 2017. p. 3. 6. KINGMA, Diederik P.; BA, Jimmy. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.

Figures

Figure 1. The network architecture. The network was composed of 6 fully connected layers with dense connection between layer 1 and layer 3. The network predicted 8 MT intensities from 4 MT intensities.

Figure 2. The PSNR of results with different frequency choices. 4 frequency set were compared; 1) 2 kHz and 25 kHz for both FAs, 2) 2 kHz and 25 kHz for 30 FA and 9 kHz and 25 kHz for 75 FA, 3) 9 kHz and 25 kHz for 30 FA and 2 kHz and 25 kHz for 75 FA and 4) 9 kHz and 25 kHz for both FAs. Bars with different letters represent statistically significant difference (p<0.05). Paired t-test was performed to compare the results obtained in the same qMT method.

Figure 3. The output of the network from the conventional and inter-slice qMT data from a representative subject. Ground truth (GT), network output and the difference image with scale 10 are displayed for each frequency and MT saturation flip angle.

Figure 4. The qMT maps and difference images with scale 2 from the same subject as that in Figure 3. The qMT fitting was performed for data from the deep learning networks and for data from the inter-slice approach (no deep learning) with 4, 6, 8, 10 and 12 (GT) MT images.

Figure 5. Comparison of qMT results. The proposed network used 4 real MT points as input and was trained to estimate 8 MT points as output, and the combined 12 points were used to generate qMT results. For comparison, qMT fitting was performed in the conventional way with no deep learning based only on 4, 6, 8 and 10 MT points. qMT fitting in the conventional way with the full 12 MT points was used as ground truth. Bars with different letters represent statistically significant difference (p<0.05). Paired t-test was performed to compare the results obtained in the same qMT method.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
3982