Introduction

Spontaneous mechanical activities in musculature (SMAM) visible in DWI have shown a strong correlation to myoelectric signals acquired by simultaneous electromyography (sEMG).^1,2 Besides spontaneous activities, other muscular contractions are also leading to signal voids in DWI, which are not distinguishable from SMAMs. However, contraction patterns are visible and well recognizable in sEMG (Figure 1). In order to investigate significant correlation and differences regarding SMAMs between healthy and non-healthy subjects and to enable cohort studies, reliable assessment and quantification of SMAMs is required. Due to this, non-spontaneous muscle contractions visible in DWI have to be excluded before data analysis. Therefore, contraction patterns have to be identified in sEMG data to reject DWI, which were affected by these non-spontaneous muscle contractions. A retrospective method based on automatic sEMG classification and DWI rejection to achieve reliable activity assessment is proposed.

Methods

Three healthy volunteers (age: 41±17, BMI: 28.5±4.2) were examined on a 3 T MR scanner (MAGNETOM Prisma^fit, Siemens Healthcare, Erlangen, Germany) with a 15-channel Tx/Rx-coil and concurrent recording of sEMG. DWI of the right calf was conducted under three different muscular contraction states: resting, active isometric contraction and active movement. Measurement parameters of DWI: prototype stimulated-echo echo-planar imaging sequence (Siemens Healthcare GmbH), matrix-size: 80 x 80, FOV: 192 x 192 mm², TE: 26 ms, TR: 500 ms, GRAPPA: 2, BW: 2015 Hz/px; slice-thickness: 6 mm, diffusion-sensitizing time Δ: 157 ms, b-value: 100 s/mm². sEMG measurements were concurrently recorded with a MR-compatible system (BrainAmp ExG MR, Brain Products GmbH, Gilching, Germany) at the DWI slice location with 5 kHz sampling rate and four channels. Each muscular contraction state was recorded over 250 s (500 repetitions in DWI). Data of four subjects, which were previously measured concurrently to DWI, were additionally taken to provide sufficient amount of training data regarding spontaneous events. sEMG classification: Framework for retrospective muscle contraction classification and data rejection is depicted in Figure 2 showing the sEMG classification pipeline and connection to DWI. First, simultaneous sEMG datasets were artifact-corrected and filtered.^2-6 All sEMG signals higher than 10 µV were split into patches and subsequently processed by a two-class Support-Vector-Machine (SVM) to detect muscular contraction. Afterwards, all patches containing sEMG activities were processed by a four-class SVM for classification with features from multiple patch-sizes. Two different groups of feature sets were employed on each patch-size step: 40 time-domain^7-11 and 31 time-frequency-domain features^7,12-13. Time-domain features were also calculated from raw sEMG signal patches to address the issue of MR residuals from MR gradients. Feature reduction was performed by classical principal component analysis. Classification was based on one-vs-one soft-margin multi-class SVM with radial-basis function utilizing LIBSVM¹⁴. Four classes were investigated: active isometric contraction (1), active movement (2), spontaneous activity (3) and MR residual artifact (4). 4680 signal patches (basic-size: 30 ms) (class: 1/2/3/4; 1825/1097/617/1141) were manually labeled by a human observer. Hyperparameters of SVM (γ,C) were optimized via cross-validation. Evaluation: F1-measure of sEMG classification was determined as a function of the number of principal components by 7-fold CV (each subject was taken once as test data). To investigate the impact of different contraction patterns onto spontaneous activity maps (event count maps – ECM), three composed datasets (one for each volunteer) were evaluated. Each dataset contained 500 DWI with following muscle contraction states: 70 % resting, 20 % isometric contraction and 10 % active movement. Thus, each dataset represents a clinical feasible acquisition. DWI were processed by self-written code in MATLAB (The MathWorks, Inc., Natick, MA, USA).¹⁵ DWI classified as "movement" or "isometric contraction" were discarded from ECM calculation to compare activity maps with and without rejection of contraction-corrupted images.

Results & Discussion

In Figure 3, result of sEMG classification is shown. F1-measure of 95.49 % was achieved with 27 principle components and four hierarchical features patches (size: 30/60/120/240 ms). Thus a reliable identification of contraction-corrupted DWI is ensured. Comparing ECM with and without data rejection reveals clear differences (Figure 4). As shown in Figure 4/5, few images containing active movement lead to overestimation of spontaneous activity. Number of DWI with signal voids were reduced from 75, 44 to 29 in average by rejection of contraction-corrupted DWI.

Conclusion

Reliable classification of sEMG contraction patterns was shown. Moreover, reliable assessment of overall spontaneous activity and improved activity maps were investigated by sEMG-based signal classification and contraction pattern analysis. Rejection of DWI with these muscle contractions, which are not from spontaneous nature, is highly recommended before data analysis to yield trustworthy data in healthy and non-healthy subjects.

Acknowledgements

We thank Stemmer, A., Siemens Healthcare GmbH, and Ramos-Murguialday, A., Institute for Medical Psychology and Behavioural Neurobiology, University of Tübingen, for their valuable technical support on this project.