Introduction

Diffusion-weighted magnetic resonance imaging (DW-MRI) is sensitive to record incoherent muscular motion like small spontaneous mechanical activities in musculature (SMAM)¹. Simultaneous recordings of DWI and surface electromyography (sEMG) have shown high correlation between both modalities.² Furthermore, a high variation of the visualization capability was revealed, depending on MR sequence parameters.³ Therefore, improved detection and visualization of muscular motion was investigated by sEMG-triggered MR-DWI in previous studies^4,5. A tradeoff between average systemic delay between onset of sEMG activity and MR-DWI⁴, and computational complexity has to be considered⁵. In this work, an improved system combining a short systemic delay with a highly adaptable noise-robust detection by neural network architectures is investigated to overcome limitations of previous works^4,5. Moreover, preliminary results regarding the influence of the trigger time delay on the visualization of spontaneous activities and active muscular contractions are presented.

Methods

Model-based event detection and MR system triggering is achieved by adaption of a sequence controller⁵ for fast signal processing. The system is depicted in Fig. 1. Sequence Controller: The sEMG signal is sampled with f_s = 1 kHz and processed galvanically decoupled on a microcontroller system (ARM Cortex-M3). For pre-processing the sEMG signal is filtered (f_band-pass=20-500 Hz and f_notch=45-55 Hz) and rectified before it is sent to the host system for storage and visualization and to the detection unit by serial communication. The sequence controller triggers the MR system by a fiber optical connection after receiving information from the detection unit. A flexible delay between sEMG onset and start of DWI measurement can be set to record motion in different states. The acquisition scheme is illustrated in Fig. 2. Detection unit: A single-board computer (Raspberry Pi Foundation, Cambridge, UK) is utilized for sEMG signal classification. For real-time detection, the inference time is reduced by a Coral Accelerator with an Edge Tensor Processing Unit (TPU) (Google LLC, Mountain View, CA, USA) and quantization-aware training⁶. Therefore, the neural network architecture is restricted to layer types which are supported by the accelerator device: Multiple layers of strided 1D convolution layers (encoder structure) with batch-normalization⁷ and rectified linear unit (ReLU) activation was utilized with a dense layer as output (softmax activation). Classification output is: "sEMG activity"/"no sEMG activity" (corresponding to a "trigger out" or "no trigger out" to the sequence controller). Training Data: The training procedure is outlined in Fig. 3. sEMG signals for training were acquired during simultaneous DW-MRI with an MR-compatible amplifier (BrainAmp ExG MR, Brain Products GmbH, Gilching, Germany) from 10 healthy subjects (age: 34±13 years, gender: 8m/2f) from previous work². sEMG signals were semi-automatically annotated.⁸ For simulation of a sEMG stream, a pre-trained signal generator⁹ (generative adversarial network^10,11 based on long short-term memory cells^12,13) was utilized to generate sEMG samples which are superimposed on resting sEMG signals recorded inside the MR system. For a more robust detection, data was augmented by additive white Gaussian noise and harmonic distortions (sinusoidal signals from 1-500 Hz). After offline training, the trained model (TensorFlow Lite, Google LLC) is transferred to the detection unit. System evaluation: sEMG signals were derived from the skeletal musculature by MR-compatible Ag/AgCL electrodes (BrainProducts GmbH, Gilching, Germany) during MR-DWI (3T MAGNETOM Prisma^fit, Siemens Healthcare, Erlangen, Germany) measurements with a 15 ch. Tx/Rx knee coil from 3 healthy subjects (age: 42±21, gender: male) with following parameterization: TE: 36 ms, b-value: 100 s/mm², 6/8 readout, matrix: 64-80x64-80, no fat-suppression, spin echo or stimulated echo excitation (diffusion-sensitive time: Δ = 145 ms).

Results & Discussion

An average inference time on the TPU of 0.4 ms was measured with a test-accuracy of 71.7 % on noisy data. Time between sEMG onset and MR trigger output was 5.4±2.8 ms in average leading to an overall minimum systemic delay of 12.6±4.5 ms due to the necessary trigger pulse length of around 10 ms which is much lower or comparable to previous works^4,5. Non-triggered MR-DWI, MR-DWI triggered by spontaneous activity and MR-DWI triggered by active movement are exemplarily depicted in Fig. 4. Clear differences are visible in the near-surface region of the sEMG electrode. Recording of active movements can be prevented by MR acquisition after muscular contraction (prolonged trigger time delay). Fig. 5 shows the influence of the trigger time delay on spontaneous activities. It is shown that with a larger trigger time delay the contraction patterns of spontaneous activities are almost fully relaxed after 175 ms.

Conclusion

Robust sEMG activity detection enables studies of small muscular movements or contraction pattern analysis by triggered MR-DWI measurements. Further work should investigate systematically the influence of the neural network architecture, e. g. sequence length and types of layers (dilated temporal convolutional networks), on the detection capability and robustness regarding signal distortions as well as the effect of the trigger time delay on the muscular contraction pattern. Furthermore, the real-time classification can improve the quantitative assessment of spontaneous activities by discarding time periods of active movement¹⁴.

Acknowledgements

This work was supported and funded by the German Research Foundation (DFG) under Grants SCHI 498/11‐1 and YA 28/16‐1. We thank Feiweier, T., Siemens Healthcare, Erlangen, Germany for his valuable technical support on this project.

References

[1]: Steidle G, Schick F. Addressing spontaneous signal voids in repetitive single-shot DWI of musculature: spatial and temporal patterns in the calves of healthy volunteers and consideration of unintended muscle activities as underlying mechanism. NMR Biomed 28(7):801-10, 2015.
[2]: Schwartz M, Steidle G, Martirosian P, Ramos-Murguialday A, Preißl H, Stemmer A, Yang B, Schick F. Spontaneous Mechanical and Electrical Activities of Human Calf Musculature at Rest Assessed by Repetitive Single-Shot Diffusion-Weighted MRI and Simultaneous Surface-Electromyography. Magn Reson Med 79(5):2784-94, 2018.
[3]: Schwartz M, Steidle G, Martirosian P, Ramos-Murguialday A, Stemmer A, Yang B, Schick F. Estimation of the Sensitivity Characteristics and Detection Capability of Diffusion-Weighted MR Sequences in Imaging Spontaneous Mechanical Activity in Musculature. Proc. ISMRM, Honolulu, USA, 2017.
[4]: Schwartz M, Martirosian P, Steidle G, Erb M, Yang B, Schick F. Feasibility of real-time surface electromyography-triggered diffusion-weighted imaging for prospective imaging of spontaneous unintentional focal muscular motion in human calf musculature. Proc. ESMRMB, Barcelona, Spain, 2017.
[5]: Schwartz M, Martirosian P, Yang B, Schick F. A Surface Electromyography-Driven Magnetic Resonance Sequence Controller for Real-Time Myoelectric Triggered Imaging. Proc. IEEE EMBC, Honolulu, HI, USA, 2018.
[6]: Jacob B, Kligys S, Chen B, Zhu M, Tang M, Howard A, Adam H, Kalenichenko D. Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference. arXiv:1712.05877, 2017.
[7]: Ioffe S, Szegedy C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. arXiv:1502.03167, 2015.
[8]: Schwartz M, Steidle G, Martirosian P, Ramos-Murguialday A, Yang B, Schick F. Spontaneous activity in sEMG in relation to signal voids in repetitive single-shot DW-MRI of the human lower leg: Comparison of measurements under varying conditions. Proc. ESMRMB, Vienna, Austria, 2016.
[9]: Schwartz M, Reuter M, Yang B, Schick F. Monte-Carlo Analysis for Quality Estimation of Gradient Correction Algorithms in Simultaneous Surface EMG-MRI Measurements using Signal Synthesis and Class Probability. Proc. IEEE EMBC, Berlin, Germany, 2019.
[10]: Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S,Courville A, Bengio Y. Generative Adversarial Nets. NeurIPS, 2014.
[11]: Mirza M, Osindero S. Conditional Generative Adversarial Nets. arXiv:1411.1784, 2014.
[12]: Hochreiter S, Schmidhuber J. Long Short-Term Memory. Neural Computation 9(8):1735-80, 1997.
[13]: Graves S. Generating Sequences With Recurrent Neural Networks. arXiv:1308.0850, 2013.
[14]: Schwartz M, Küstner T, Martirosian P, Machann J, Steidle G, Yang B, Schick F. Robust Quantification of Spontaneous Muscular Activities by Simultaneous Interpretation of sEMG Data. Proc. ESMRMB, Rotterdam, Netherlands, 2019.