0496

Accurate and Fast Reconstruction of Magnetic Resonance Fingerprinting for Water-Fat Separation in Neuromuscular Diseases
Fabian Balsiger1,2,3, Olivier Scheidegger4,5, Pierre G Carlier2,3, Mauricio Reyes1, and Benjamin Marty2,3

1Institute for Surgical Technology and Biomechanics, University of Bern, Bern, Switzerland, 2NMR Laboratory, Institute of Myology, Paris, France, 3NMR Laboratory, CEA, DRF, IBFJ, MIRCen, Paris, France, 4Department of Neurology, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland, 5Support Center for Advanced Neuroimaging (SCAN), Institute for Diagnostic and Interventional Neuroradiology, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland

Synopsis

The MR fingerprinting sequence MRF-WF is tailored for water and fat separation imaging for neuromuscular diseases (NMDs). Currently, the adoption of MRF-WF in the clinics is hindered by the long MR map reconstruction time of four hours per image slice. We propose a spatiotemporal convolutional neural network (CNN) to reconstruct the MR maps. We show that our CNN is robust to a highly heterogeneous dataset including patients with various NMDs. The method might be a possible solution for clinical application of MRF-WF for NMDs due to the significantly lowered reconstruction time.

Introduction

Quantitative MRI is increasingly used for diagnosing and monitoring neuromuscular diseases (NMDs). While intramuscular fat fraction (FF) quantifies degenerative processes1, muscle water T2 and water T1 relaxation times (T1H2O) can be used as non-specific biomarkers of disease activity1,2. However, quantifying several MR parameters of interest is slow and hinders the wide adoption of quantitative measures for NMDs into the clinics. Recently, MR fingerprinting (MRF) was proposed for fast and simultaneous MR parameter mapping3. For NMDs, we introduced MRF with water and fat separation4 (MRF-WF), which allows for simultaneous T1H2O and FF estimation. While the short acquisition time of MRF-WF with 10 seconds per image slice is compatible with clinical use, the map reconstruction requires approximately four hours per image slice, which make MRF-WF unfeasible for clinical use. Recent work aiming at accelerating MRF reconstruction included deep learning methods with promising results5,6,7,8. We present a convolutional neural network (CNN) for MRF-WF reconstruction that can handle two grades of undersampling levels. We evaluated its robustness to a heterogeneous dataset consisting of patients imaged in legs and thighs, with various NMDs and high variability of MR map appearances.

Methods

Data: All experiments were performed on a 3 T PrismaFit scanner (Siemens Healthineers, Germany). The MRF-WF sequence was acquired at the legs and thighs levels in 95 patients with various NMDs, resulting in a highly heterogeneous dataset (Table 1). The acquisition consisted of a non-selective inversion followed by a 1400 radial spokes FLASH echo train (golden angle scheme) with varying TE, TR, and nominal flip angle (FOV=350×350mm2, voxel size=1.0×1.0×8.0mm3, 5 slices, Tacq=50s). We reconstructed two image space series each consisting of 175 temporal images from the raw data: i) an image series reconstructed by view sharing (k-space weighted image contrast filter with 55 spokes) and compressed sensing with total variation9,10 (MRF-WFVS+CS), and ii) a highly undersampled image series with NUFFT using 8 spokes per image (MRF-WFUS). For each scan, T1H2O, FF, and transmit field efficacy (B1) reference maps were reconstructed from MRF-WFVS+CS using dictionary matching4.

CNN reconstruction: We built on our previous work5 that leverages the spatiotemporal relationship between neighboring fingerprints. Therefore, we reconstructed the maps patch-wise, i.e. input to our CNN were 2-D complex-valued image patches of size 20×20×175 and output of the CNN were three real-valued image patches of size 16×16, corresponding to the T1H2O, FF, and B1 maps (Figure 1A). The CNN itself consisted of four temporal convolutions and three spatial convolutions with varying receptive fields (Figure 1B). Each convolution applied a 2-D convolution followed by a rectified linear activation function (except the last convolution with linear activation). The CNN can handle both MRF-WFVS+CS and MRF-WFUS as input without altering any parameters of our network (Figure 1C). We trained the CNN for 100 epochs with a batch size of 50 randomly selected patches and used the Adam optimizer to minimize a mean squared error (MSE) loss with a learning rate of 0.001. The CNN was implemented with TensorFlow 1.10.0 (Python 3.6.5).

Evaluation: We split the dataset five times randomly into training/validation/testing sets (n=55/20/20) without stratifying regarding leg, thigh, and NMD to show the robustness of our method to the heterogeneous dataset. The results on the five testing sets were aggregated, and we calculated the coefficient of determination (R2), MSE, and normalized root mean squared error (NRMSE) between the CNN’s predictions and the reference maps.

Results

Qualitatively, the reconstructed maps had a high similarity to the reference maps (Figure 2). Box-and-whisker plots of the R2 for the MRF-WFUS and MRF-WFVS+CS reconstruction are shown in Figure 3. Additional descriptive statistics are summarized in Table 2. The reconstruction of MRF-WFVS+CS achieved high correlations with mean R2s above 0.90 for all maps. The reconstruction of MRF-WFUS resulted in mean R2s above 0.77. The total map reconstruction times were approximately 2 hours and 10 minutes for MRF-WFVS+CS and MRF-WFUS, respectively.

Discussion & Conclusion

Our results suggest that our spatiotemporal CNN is able to learn the mapping of neighboring fingerprints from different undersampling levels to quantitative MR parameters at interest. While the FF map could be reconstructed with similar accuracy independent of the image series (MRF-WFVS+CS and MRF-WFUS), the T1H2O and B1 maps were less accurate reconstructed from MRF-WFUS. Further work will focus on improving the accuracy for MRF-WFUS. In conclusion, we proposed a CNN that can reconstruct MR maps from two grades of undersampled MRF data, and that is robust to a highly heterogeneous dataset. The method might be a possible solution for clinical application of MRF-WF for NMDs due to the significantly lowered reconstruction time.

Acknowledgements

No acknowledgement found.

References

1. Carlier PG, Marty B, Scheidegger O, et al. Skeletal Muscle Quantitative Nuclear Magnetic Resonance Imaging and Spectroscopy as an Outcome Measure for Clinical Trials. Journal of Neuromuscular Diseases. 2016;3(1):1-28. https://doi.org/10.3233/JND-160145

2. Marty B, Coppa B, Baudin P-Y, et al. Physiological and pathological skeletal muscle water T1 changes quantified using fast radial T1 mapping. Magnetic Resonance Materials in Physics, Biology and Medicine. 2017;30:63. https://doi.org/10.1007/s10334-017-0632-1

3. Ma D, Gulani V, Seiberlich N, et al. Magnetic resonance fingerprinting. Nature. 2013;495(7440):187-192. https://doi.org/10.1038/nature11971

4. Marty B and Carlier PG. Quantification of water T1 and fat fraction in skeletal muscle tissue using an optimal MR fingerprinting radial sequence (MRF-WF). 26th Proceedings of the International Society for Magnetic Resonance in Medicine. 2018.

5. Balsiger F, Shridhar Konar A, Chikop S, et al. Magnetic Resonance Fingerprinting Reconstruction via Spatiotemporal Convolutional Neural Networks. In Machine Learning for Medical Image Reconstruction. 2018;11074:39-46. Cham: Springer. https://doi.org/10.1007/978-3-030-00129-2_5

6. Fang Z, Chen Y, Liu M, et al. Deep Learning for Fast and Spatially-Constrained Tissue Quantification from Highly-Undersampled Data in Magnetic Resonance Fingerprinting (MRF). In Machine Learning in Medical Imaging 2018;11046:398-405. Cham: Springer. https://doi.org/10.1007/978-3-030-00919-9_46

7. Cohen O, Zhu B, Rosen MS. MR fingerprinting Deep RecOnstruction NEtwork (DRONE). Magnetic Resonance in Medicine, 2018;80(3):885–894. https://doi.org/10.1002/mrm.27198

8. Hoppe E, Körzdörfer G, Nittka M, et al. Deep Learning for Magnetic Resonance Fingerprinting: Accelerating the Reconstruction of Quantitative Relaxation Maps. 26th Proceedings of the International Society for Magnetic Resonance in Medicine. 2018.

9. Marty B, Coppa B, Carlier PG. Monitoring skeletal muscle chronic fatty degenerations with fast T1-mapping. European Radiology. 2018;28(11):4662-4668. https://doi.org/10.1007/s00330-018-5433-z

10. Marty B, Coppa B, Carlier PG. Fast, Precise, and Accurate Myocardial T1 Mapping Using a Radial MOLLI Sequence With FLASH Readout. Magnetic Resonance in Medicine. 2018;79(3):1387-1398. https://doi.org/10.1002/mrm.2679

Figures

Table 1: Summary of our highly heterogeneous dataset with patients affected by various neuromuscular diseases (NMDs), and a wide range of patient’s ages. The dataset is mixed with images of the leg and thigh. Ages are given in mean ± standard deviation.

Figure 1: Overview of the proposed method. A) The method reconstructs the T1H2O, FF, and B1 maps patch-wise from an image slice of either MRF-WFVS+CS or MRF-WFUS (MRF-WFVS+CS shown). B) The architecture of our spatiotemporal CNN. The number of channels are denoted on the top of the bars (T=175 and M=3), and the spatial image patch size is provided at the lower left edge of the bars (conv: convolution, ReLU: rectified linear unit). C) The image quality differs considerably between MRF-WFVS+CS and MRF-WFUS.

Figure 2: Reconstruction of a 69 years old female patient with inclusion body myositis. The reference maps, the reconstructions from MRF-WFVS+CS and MRF-WFUS, and the reconstruction errors (reference minus prediction) are shown for the T1H2O, FF, and B1 maps. Background voxels were excluded for the FF and B1 maps using a mask segmented on an anatomical image (out-of-phase). The mask for the T1H2O map was further processed to exclude voxels with a FF higher than 0.65.

Figure 3: Box-and-whisker plots of the coefficient of determination (R2) for the MRF-WFVS+CS and MRF-WFUS reconstruction of the A) T1H2O, B) FF, and C) B1 maps. Background voxels were excluded for the FF and B1 maps using a mask segmented on an anatomical image (out-of-phase). The mask for the T1H2O map was further processed to exclude voxels with a FF higher than 0.65.

Table 2: Mean and standard deviation of the coefficient of determination (R2), mean squared error (MSE), and normalized root mean squared error (NRMSE) for the T1H2O, FF, and B1 map reconstructions from MRF-WFVS+CS and MRF-WFUS. Background voxels were excluded for the FF and B1 maps using a mask segmented on an anatomical image (out-of-phase). The mask for the T1H2O map was further processed to exclude voxels with a FF higher than 0.65.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
0496