Advertisement

Automatic analysis of ultrasound shear-wave elastography in skeletal muscle without non-contractile tissue contamination

  • Ellenor Brown
  • Yasuhide Yoshitake
  • Minoru Shinohara
  • Jun Ueda
Regular Paper

Abstract

Current analysis methods for obtaining mean shear modulus of skeletal muscles with ultrasound shear-wave elastography are limited by contamination with non-contractile tissues and manual operation of video processing. In this work, we develop new ultrasound image processing methods to assess muscle activities. We build upon previous research by using a 6-DOF robotic manipulator system and indirectly quantifying extensor carpi ulnaris (ECU) and triceps brachii longus (TRI) muscle elasticity during loading using ultrasound shear-wave shear modulus elastography. The purposes of this study were to (1) develop an automatic image-processing algorithm for removing non-contractile tissues from muscle elastography videos and (2) understand the effect of the removal on comparison of mean shear modulus of muscles across static motor tasks with variable muscle loadings in healthy humans. The developed algorithm with optimized clustering and thresholding identified and removed non-contractile tissues from muscle elastography videos with > 90% accuracy in arm muscles, causing reductions in the spatial variability of shear modulus data within the region of interest in healthy young adults. Removal of non-contractile tissues can alter the mean shear modulus of the muscles and influenced task comparisons by substantially altering the ranking of tasks by mean shear modulus.

Keywords

Shear modulus Shear wave elastography Image segmentation Skeletal muscle Human–robot interaction 

1 Introduction

Quantitative and qualitative assessment of joint torques (e.g., manual muscle testing, grip strength) is routinely used in clinical settings as a diagnostic tool for and as a means to monitor rehabilitation from neuromuscular disease and injury (e.g., stroke, spinal cord injury). Assessment of mechanical activity in individual contracting muscles would also be informative for diagnosis because the relative contributions of individual muscles to producing joint torque combinations (i.e. muscle coordination) often appear to be abnormal in the above-mentioned clinical populations according to the electrical activity of the muscles (Clark et al. 2010; Rodriguez et al. 2013). In humans in vivo, however, mechanical output of contracting muscles cannot be quantified by direct measurement of individual muscle forces noninvasively.

As alternatives, surface electromyography, mechanomyography, and laser shear-wave elastography have been studied for estimating muscle force, but they all have various fundamental limitations. Surface electromyography measures electrical activity of muscles from skin surface, but its amplitude is not always linearly related to mechanical output of muscle (i.e. force) (Lawrence and De Luca 1983), or the linear relationship is reduced by task conditions (Bouillard et al. 2012). Mechanomyography measures small lateral vibration of muscles from skin surface, but its amplitude is not always linearly related to muscle force especially at higher contraction levels. (Beck et al. 2004; Madeleine et al. 2001; Shinohara and Søgaard 2006) Laser shear-wave elastography has recently been developed to measure propagation velocity of shear wave at skin surface overlaying an upper arm muscle (Salman et al. 2014), but its applicability to other muscles is yet to be examined. All of these recordings from the skin surface have common fundamental limitations including difficulty in obtaining reliable signals from deep muscles, variability across subjects, and susceptibility to crosstalk from neighboring muscles, particularly on body segments like the forearm where there are multiple small, closely packed, muscles (e.g. wrist muscles) with overlapping (i.e. redundant) function (Kong et al. 2010; Mogk and Keir 2003). To overcome these limitations, we are one of the first groups to demonstrate the potential of recently developed ultrasound shear-wave elastography for assessing individual muscle contraction levels (Shinohara et al. 2010). With the growing interest in the development of robotic systems for neuromotor training and rehabilitation by our and other groups (Klamroth-Marganska et al. 2014; Ueda et al. 2010; Liao et al. 2012), we envision the potential applicability of ultrasound shear-wave elastography for assessing muscle contraction intensity during rather complex interaction with a robotic system. To make it possible, we are motivated to solve essential problems of the current use of ultrasound shear-wave elastography in contracting muscles.

In ultrasound shear-wave elastography, shear modulus is calculated from the propagation velocity of shear wave inside tissues, which can be spatially variable when tissues are made of different elasticities (Bercoff et al. 2003; Bercoff et al. 2004). In its current application to muscles, the spatial average of shear modulus distribution across a region of interest (ROI) of a muscle has been determined using a commercially available ultrasound device with included software (e.g. AixPlorer system Real-time ShearWave Elastography with Q-BOX software, Supersonic Imagine, Aix-en-Provence, France) (Akagi et al. 2015; Bouillard et al. 2011; Kot et al. 2012; Yoshitake et al. 2014). Thus obtained mean shear modulus of a large muscle is shown to be more linearly related to joint torque over a wide range of contraction level with less variability compared with surface electromyography (Bouillard et al. 2012; Yoshitake et al. 2014) and not susceptible to crosstalk (Nordez and Hug 2010). The problem is that a simple spatial average of shear modulus distribution across an ROI in a muscle is not always appropriate for assessing contraction level because it can include not only shear modulus of contractile tissues (i.e. muscle fibers) that produce active force with contraction, but also non-contractile tissues (NCT, i.e., skin, adipose tissue, and connective tissue) that can be distributed within the contractile tissue region. While connective tissue can contribute to passive muscle force production, changes in muscle force over time or compared across tasks are expected to be directly related to the activity of contractile tissue. Therefore, the selection of an ROI should allow for avoidance of NCT while maximizing the inclusion of contractile tissue, which may be limited by the size of the muscle and the distribution of NCT. Still, the impact of NCT inclusion in the image analysis on the assessment of muscle contraction intensity is unknown.

Additionally, the above mentioned determination of mean shear modulus is often manually operated image by image with multiple steps, which is time consuming and prone to error and variability due to subjective placement of ROI in examining multiple images across trials, tasks, and individuals. Automated image processing algorithms have been developed for medical images, but many are specialized for tissues or organs other than skeletal muscle, employ methods that are not appropriate for NCT detection, or introduce more complexity than our study requires. These problems will need to be resolved for enhancing the accuracy and convenience of using ultrasound shear-wave elastography for assessing contracting muscles.

The purposes of this study were (1) to develop an automated analysis method for determining mean shear modulus of muscles without the contamination of NCT in ultrasound shear-wave elastography and (2) to understand the effect of NCT on comparison of mean shear modulus of muscles across tasks with variable muscle loadings when interacting with a robotic system. For the first purpose, we hypothesized that the implementation of an optimized clustering and thresholding algorithm for ultrasound B-mode images could automatically identify NCT and be used to successfully remove shear modulus values of NCT from spatial averaging of muscle shear modulus. Since elasticity of NCT is distinct from contractile tissues (Kot et al. 2012; Lacourpaille et al. 2012), the successful removal of these “outliers” is expected to reduce the variability of shear modulus value within ROI. For the second purpose, we hypothesized that removal of NCT contamination alters mean shear modulus of muscle, and as a consequence, impacts the rank order of motor tasks that is determined from mean shear modulus during the interaction with a robotic system.

2 Materials and methods

2.1 Subjects

Eleven male subjects [age: 21.3 ± 0.9 years (mean ± SD)] provided informed consent according to the procedures approved by the ethics committees of the National Institute of Fitness and Sports in Japan, the Nara Institute of Science and Technology in Japan, and the Georgia Institute of Technology in Atlanta, GA, USA. All subjects were college athletes who were physically and cognitively healthy with no history of neuromuscular disorder. Male athletes were chosen for their low subcutaneous fat and for their expected ability to control and stabilize their muscle activity. These attributes allow for optimal ultrasound image acquisition. Further, the homogeneity of this subject group was expected to reduce measurement- and subject-dependent variability in the data, thus allowing us to use a smaller number of subjects for statistical analysis and to focus on development of the algorithm and assessment of the impact of NCT inclusion efficiently.

2.2 Motor task

The motor task for the subjects was to maintain their upper limb posture against various combinations of force and torque loads applied by a robotic manipulator. A task against various forces and torques was utilized to obtain data in which healthy individuals may use various “non-standard” combinations of muscle activation as in clinical populations. Subjects were seated and grasped the handle of the robotic manipulator with the right hand. The starting position of the robotic manipulator handle was adjusted for each subject such that the shoulder joint was abducted to approximately 10°, the elbow joint was flexed to 90°, and the wrist joint was in a neutral position (standard posture). The elbow joint angle was monitored with an electro-goniometer (SG110, Biometrics Ltd, Newport, UK) and displayed on an analogue output unit (ADU301, Biometrics Ltd, Newport, UK) to maintain the right angle throughout the measurements. While watching their hand and arm, subjects were instructed to contract their upper-limb muscles to maintain the standard posture against the complex force and torque loads applied by the robot manipulator. This experimental setup can be found in a picture that was taken during a pilot study (Fig. 1).
Fig. 1

A photo that includes the current experimental setup, taken during a pilot experiment. Axes of loading for robotic manipulator are: positive x, y and z axes point toward the subject, to the right of the subject, and upward, respectively. Note that the subject is being prepared for an upcoming trial, thus the elbow has not yet been adjusted to 90° of flexion. In addition, pilot EMG signals were explored for a purpose not addressed in this paper but were deemed unreliable due to motor noise from the robot

The target muscles in this study were extensor carpi ulnaris (ECU) and triceps brachii longus (TRI). ECU was chosen as a representative forearm muscle to illustrate the use of elastography in a body segment prone to EMG crosstalk. TRI was chosen to represent larger muscles where reliability of EMG may often be degraded with relatively thick subcutaneous adipose tissue. ECU plays a role in wrist extension and ulnar deviation while TRI plays a role in elbow and shoulder extension. We were also interested in testing and exploiting the ability of the system to control bi-articular muscles and to deal with high muscle redundancy. Upper arm muscles are larger and less redundant while forearm muscles are smaller and more redundant. Increased muscle redundancy in the forearm presented a greater challenge for individual muscle control. Both TRI and ECU are bi-articular muscles.

Prior to human subject testing, the computational model was used to plan the experimental loading tasks. Thirteen different types of muscle loadings were selected for the robotic manipulator such that the expected force/torque production (i.e. contraction level) in ECU and TRI would vary substantially across tasks and between muscles for the selected arm posture. The forces and torques for each loading condition are listed in Table 1. Tasks 1, 2 and 3 are linear loading conditions and were used to test linearity of muscle activity captured with elasticity data. Details of the loadings were determined by modifying our computational muscle model that contains 51 muscles across 12 joints of the human right arm and torso (Ueda et al. 2010; Gallagher et al. 2013) for the robotic device. Nonetheless, with the abundancy of muscle contraction patterns for accomplishing the same end-point task performance (Bernstein, N.A.: The co-ordination and regulation of movements 1967; Latash 2012), substantial variability of contraction level was anticipated in ECU and TRI across subjects. Each subject performed three 15-s trials of each loading task for each muscle. Subjects were instructed to use the same muscle activation strategies across trials for the same loading configuration. All subjects were able to resist the applied loads with minimal displacement of the arm and hand.
Table 1

A list of forces (Fx, Fy, Fz) and torques (Mx, My, Mz) applied by the robot for each loading condition

Task #

Fx (N)

Fy (N)

Fz (N)

Mx (N m)

My (N m)

Mz (N m)

1

0

0

20

0

− 0.5

0

2

0

0

10

0

− 0.25

0

3

0

0

30

0

− 0.75

0

4

0

0

22.5

− 0.9

0.1

0

5

0

0

10.8

2.3

− 3

0

6

0

0

8.3

0.8

− 0.9

0

7

0

0

42.7

− 2.9

0.3

0

8

0

0

1.7

1.7

− 3.3

0

9

− 5.7

3.6

2.4

1.5

− 3.4

0

10

30.7

− 2.6

31.7

− 0.6

0

0

11

− 0.8

0

8.6

0.7

− 0.9

0

12

− 4.2

2.8

11.2

2.1

− 3.1

0

13

− 7.8

0.6

− 19.7

0.3

− 0.1

0

2.3 Robotic manipulator and loading conditions

The robotic manipulator used in this study was the PA10 (Mitsubishi Heavy Industries, LTD.). The PA10 has 7 degrees of freedom to apply forces and torques on the subject’s hand. A handle which included a force sensor (MICRO 5/50, BL Autotec, Ltd.) was attached to the end of the PA10. The basic idea of the control algorithm for the manipulator was to control the position and rotation of the handle using force sensor feedback. Figure 2 shows the control algorithm schema and its integration with the robotic control algorithm. To start, the position of the PA10 was adjusted to accommodate the arm posture of the subject. As mentioned above, the elbow angle of each subject was set to 90° of flexion when the subject was holding the handle and their wrist was in a neutral position. The robot was set to apply the specified loading conditions for each task.
Fig. 2

Muscle and robot control schema, where ftd is desired target muscle force, Fd = desired human arm forces and torques, F is robot loading, J is the Jacobian matrix relating joint torques to end-point loading at the hand and τex = total external torque on the hand

For the experiments, a motion control board (Mitsubishi Heavy Industries) was employed as a proportional integral derivative (PID) controller for the handle position, rotation, and applied force. The handle position and rotation were transformed into the manipulator joint angles by the motion control board. The third rotation axis of the PA10 was locked in the initial angle to avoid redundancy. The target position sent to the motion control board was obtained by summing the neutral position and a differential position calculated from the force sensor data. The differential position was obtained by taking the difference between the target force and the force sensor reading. The handle movement increased and decreased the applied load to maintain the desired loading as the subject performed each task.

The manipulator was programmed to apply force and torque loads up to six degrees of freedom (Fx, Fy, Fz, Mx, My, Mz) at the handgrip. However, subjects performed tasks of 3DOF (Fz, Mx, My) and 5DOF (Fx, Fy, Fz, Mx, My). The former simulates conditions achievable through gravity induced loading. Mz was excluded in general due to the subjects’ difficulty with resisting this rotation without the handgrip sliding in the palm.

A computational muscle force model of the right arm was used to select the loading conditions with the intent to induce a range of activation levels for the target muscles. This computational model has been implemented and published (Gallagher et al. 2013), but was modified for the robotic manipulator used in this study. An overview of the computation model is presented in the Appendix.

2.4 Ultrasound data recording

Ultrasound B-mode and elastography videos were obtained concurrently during each task using a commercially-available ultrasound shear-wave elastography system (AixPlorer Ver. 4.0, Supersonic Imagine, Aix-en-Provence, France) with the musculoskeletal preset. The ultrasound system provides a repeatable and reliable method of quantifying mean shear modulus of muscles at various contraction levels, showing linear increases in mean shear modulus with increasing muscle activity (Yoshitake et al. 2014; Koo et al. 2013). In this system, shear modulus is calculated from the shear-wave propagation velocity in the direction of the longitudinal axis of a probe (4–15 MHz, SL15-4, linear-array, 50-mm wide, Supersonic Imagine, Aix-en-Provence, France). In this experiment, the probe was held over the target muscle belly and aligned with the longitudinal axis of the muscle. Care was taken not to press or deform the muscle with the probe during measurements. ECU and TRI could not be monitored simultaneously as only one ultrasound device was available. Therefore, loading configurations were performed to record one muscle and later repeated to record the other. B-mode and elastography data from each trial and muscle were saved as video (1 frame/s). Representative B-mode and elastography images of ECU and TRI are shown in Fig. 2a, b, respectively, to show the differing architectures of the muscles. A rectangular area within the muscle, outlined in the B-mode images, was selected as the recording area for shear modulus data during the experiment. The recording area for shear modulus was chosen to span the thickness of the muscle belly in ECU or to capture data from a central location in the muscle belly of TRI.

2.5 Data analysis

Ultrasound B-mode and elastography data were processed offline in MATLAB (Mathworks, Natick, MA,USA). In each trial, approximately 12 s of continuous video with consistent images were extracted for data analysis. Each video frame included two grayscale (B-mode) images of the underlying arm anatomy. One featured elastography, showing a color-coded rectangular overlay representing the calculated Young’s modulus (Fig. 3, upper images). The other was a copy of the B-mode image with only a colored rectangular outline corresponding to where Young’s modulus was calculated (Fig. 3, lower images). The entire rectangular measurement area was used as the ROI from which data were processed. The dimension of the ROI was approximately 1.0 × 1.5 cm for ECU and 1.5 × 1.5 cm for TRI. The system calculated Young’s modulus from measured shear wave velocity with the assumption that the measured substrate was isotropic. As muscle is anisotropic, the calculated Young’s modulus data were transformed to shear modulus by dividing the Young’s modulus values by three as in our previous studies (Yoshitake et al. 2014; Maher et al. 2013; Taniguchi et al. 2015). For each trial, the scale of the color mapping for the modulus was adjusted to provide optimal visual resolution, i.e., large visual change of color without saturation to either end of the color scale. No trials reached the maximum measurable limit (266 kPa for shear modulus) of the ultrasound device.
Fig. 3

Representative ultrasound images of the extensor carpi ulnaris (ECU) (a) and the triceps brachii (TRI) (b). A colored box over the B-mode image (top) shows the region of interest (ROI) selected for the elastography recording. The long white bands of tissue across the middle of each image are connective tissue. The flexor digitorum profundus (FDP) is shown deep to ECU

2.6 Image processing for NCT identification

The greatest challenge of the automated ultrasound image processing was successful detection of NCT across muscles and subjects without excessive removal of contractile tissue data. A method was needed that was robust to differing NCT configurations and locations and occasional image quality issues. Several algorithms for analysis of ultrasound images exist with various applications from medical imaging (Darby et al. 2013; Graf et al. 1999; Loram et al. 2006; Peng et al. 2006; Yang et al. 2002) to food quality assessments (Brosnan and Sun 2004; Du and Sun 2004; Kim et al. 1998). These algorithms employ edge detection or image segmentation using pixel intensities, textures, or developed templates within a single image or tracked across several images in a video. The established edge detection algorithms were not appropriate for our use, as our main objective was to identify the full area of objects, i.e. non-contractile tissue segments. Further, the NCT within the ultrasound images did not have enough sharp, continuous edges for sufficient edge detection. For segmentation of the image into areas of interest, intensity values (Yang et al. 2002) and visual texture (Kim et al. 1998) have been used to distinguish between tissues. For the analysis of ultrasound images of muscle for NCT removal, the tissues of interest were distinguishable by their differing intensities, which obviated the need for the more complex texture analysis. With distinct intensity values, methods of pixel clustering by intensity for tissue identification were most appropriate. For the above-mentioned reasons, we developed an image processing algorithm that identified bands of NCT based on intensities in the B-mode ultrasound image captured along the longitudinal axis of the muscle (Fig. 4). Building upon established methods of intensity value clustering, we modified our intensity threshold and cluster shapes to best suit the band-like nature of most NCT segments.
Fig. 4

Progression of ultrasound B-mode images via image processing algorithm using an example of TRI. a Original B-mode image. b Transformed to black and white image using Otsu’s intensity threshold multiplied by X. c Segments with length below the Yth percentile removed. d Segments with eccentricity above the Zth percentile were lengthened by W%. See text for details of W, X, Y, and Z

The initial part of the algorithm included determination of intensity threshold and clustering variables for identifying continuous NCT areas from distributed echo intensities. To start, each ultrasound video was converted into its constituent frames for processing in MATLAB. In each frame, the B-mode ultrasound image without the color-coded ROI (Fig. 4a) was converted into a black and white image using the built-in MATLAB functions (“im2bw” and “graythresh”). “Graythresh” used Otsu’s method to choose an intensity threshold that minimized the intraclass variance of the grouped pixels (Otsu 1979). The intensity threshold and ultrasound image were inputs to “im2bw”, wherein the ultrasound image pixels with intensities below the threshold were shown as black and assigned the intensity value 0, while all other pixels were assigned the intensity value 1. For our analysis, the intensity threshold from “graythresh” was multiplied by X to bias the algorithm toward detecting more or less potential NCT in the initial stage (Fig. 4b).

The white pixels were further examined to find major segments of NCT. The function “regionprops” was used to find the major axis length, eccentricity, and orientation of all potential NCT segments, or 8-connected objects, within the set of pixels with intensity value 1. Ideally, segments of NCT would appear as long, narrow lines of white pixels spanning the length of the frame. Therefore, we were interested in long segments with high eccentricities. First, we applied a length threshold: segments that were below the Yth percentile of major axis lengths across all segments in the image were discarded (Fig. 4c). Next, segments with eccentricity above the Zth percentile were lengthened by W% on each end along the major axis (Fig. 4d). This lengthening was implemented for cases where long bands of NCT were artificially broken up due to inadequate image quality, low contrast, or limitations of the intensity thresholding. Lengthening was accomplished as an image dilation using a linear “morphological structuring element” or “strel” with the same orientation as the segment to be lengthened and a length equal to W% of the segment’s original major axis length. Finally, all segments were transferred to a final NCT binary mask.

This NCT detection algorithm was trained on 88 images (four from each muscle and subject) randomly selected from the obtained images by the authors to determine values for X, Y, Z, and W. Each resulting binary mask was compared to a “solution”, a manually segmented version of the test image created by the authors beforehand. Results were scored based on the percentage of true positive pixels (correctly identified as NCT out of all NCT) and a combined score of the true positives and true negative pixels (correctly identified as contractile tissues out of all contractile tissue). For a given set of variable values, the true positive, true negative, and combined score were averaged across all 88 images.

To optimize the variable set, a greedy algorithm was used to maximize the combined score while preferentially selecting variable sets with true positive values above 90%. To start, a random set of values was used as a seed. The random initial seed values for each variable were from 0.5 to 1.5 for X, 50 to 100% for Y and Z, and from 0 to 100% for W. From this seed, eight variable sets were generated by increasing and decreasing each of the four variable values separately by a small increment of 0.05, 2.5, 2.5, and 10% for X, Y, Z, and W, respectively. Thus, a total list of nine variable sets was tested initially. The next seed was determined by first finding the variable set with the highest combined score among those with true positive scores above 90%. If none of the true positive scores was above 90%, the algorithm selected the variable set with the highest combined score. If the true positive score was above 90% but the combined score was less the 5% lower than the highest combined score in the list, the variable set with the highest combined score was selected instead. To avoid infinite loops resulting from a repeated seed, a randomized process selected one value in the repeated seed, incremented or decremented the value, and then used the resulting variable set as the next seed. The algorithm continued until a seed had a true positive score above 90% and had either the highest combined score or was within 5% of the highest combined score in the current list. This greedy algorithm was used in order to explore the variable set space without having to run every possible option within an arbitrarily limited range of values. There was not expected to be a straightforward global maximum for the combined scores, so the variable set search was repeated 10 times to give a selection of variable sets and give insights on optimal ranges for each variable. Of the 10 variable sets [X Y Z W], the set with the highest combined score while having a true positive score about 90% was selected for analysis of the video data. This optimized value set was used to identify NCT in each frame of the ultrasound videos, producing a binary mask (Fig. 4d) used to remove NCT data values from the ultrasound images.

2.7 Image processing for shear modulus assessment

Shear modulus values were read from each frame with and without inclusion of areas of NCT. Detection of color-coded shear modulus data was based on finding image pixels with high variability among RGB components, given that the rest of the image is essentially in grayscale with low variability. For each colored pixel detected, the algorithm found the closest matching pixel in the color scale displayed on the screen, and this color was translated into a shear modulus value based on the maximum value listed above the color scale. The maximum value was user selected in the user interface of the ultrasound system. The scale had 220 colors, signifying values from 0 to the maximum shear modulus value for each video.

Once shear modulus was quantified for each pixel in a given frame, the spatial mean value of shear modulus data in the ROI was calculated for each frame in a trial video with NCT. The frames with the least variability (i.e. SD of the mean values) for a 3-s window were extracted from the video for further analysis, assuming the window indicated when muscle activity had settled. The spatial mean and SD of shear modulus in the ROI of these frames were determined with and without NCT for a given video. In addition, the relative amount of detected NCT area within the ROI was determined in percentage of the total ROI area. These data were averaged across the extracted frames in each trial and then across trials for each loading configuration in each muscle.

It is possible that the identified NCT contains both adipose tissue with low shear modulus and connective tissue with high shear modulus. Adipose tissue appears primarily at the subcutaneous level and connective tissue appears deeper and within the muscle. To examine the depth-related distribution of NCT removal in the ROI, the amount of NCT removed at each 10% of equally-distributed levels of depth in the image (~ 1.5 mm/level) was calculated as a percentage of all NCT removed in the ROI. Thus obtained NCT removal at each depth level was averaged across all frames, trials, and loading configurations for each subject and muscle. This analysis was performed to obtain information that may help infer the relative amount of the superficial (adipose) tissue in the NCT removal compared with deeper (intramuscular) connective tissue.

To determine the impact of NCT removal, the linearity of mean shear modulus across the linear loading conditions was quantified as Pearson’s coefficient and compared with and without NCT. In addition to analyzing linear loading, the rank order of tasks by increasing mean shear modulus was compared with and without NCT removal for each subject and muscle. Rank order was used as a means to express relative amounts of muscle activity under the complex loading configurations unique to this study. Differences in task ranking, or rank errors, were tabulated across subjects at each rank with NCT removal. For example, if a task is ranked at 1 (i.e. lowest mean shear modulus) with NCT removal for a given subject but is not ranked at 1 without NCT removal, this is counted as one rank error at rank 1. The total number of rank errors was also determined for each muscle. For each muscle, there were 13 ranks in each subject, and thus 143 rank error possibilities across 11 subjects (i.e. 11 × 13).

2.8 Statistical analysis

Spatial mean and SD of shear modulus in the ROI were tested with a two-way analysis of variance (ANOVA) with repeated measures with factors being NCT and muscle. The distribution of NCT removal in reference to depth was also tested with a two-way ANOVA with repeated measures with factors being depth and muscle. When appropriate, post hoc comparisons were performed using Bonferroni correction. The amount of NCT area was compared between muscles using a dependent t test. Pearson’s correlation analyses were performed between the NCT area and changes in spatial mean and SD of shear modulus in the ROI for each muscle independently and also for both muscles together across subjects. The alpha value was set to 0.05. P < 0.05 and P < 0.01 are noted when significant. Values are reported as mean ± SD in the text mean and mean and standard error in the figures.

3 Results

3.1 Algorithm optimization

The determination of variable sets of [X Y Z W] was crucial for identifying NCT. When only Otsu’s method was used for thresholding of the intensity values for identifying NCT in the original B-mode image, the variable set of [1.0 0% 0% 0%] for [X Y Z W] had true positive score of 66.58%, far less than 90% required for NCT, and true negative score of 95.04%. In contrast, of the 10 variable sets [X Y Z W] found in our optimization search, the set [0.80 95% 90% 10%] had the highest combined score while having a true positive score above 90% (true positive score: 90.01%; true negative score: 81.46%). Hence, the true positive score with the uses of Otsu’s method only was less than the method with additional optimization by 23.43% (i.e. 90.01–66.58%). The combined score was 161.62 with Otsu’s method only and 171.47 with the optimized value set [0.80 95% 90% 10%]. This optimized value set was used for the subsequent analysis of the video data in assessing the effects of NCT.

3.2 Spatial SD and mean of shear modulus

There were main effects of NCT and muscle on both spatial SD and mean of shear modulus in the ROI (P < 0.01 for all the main effects) (Fig. 5). With the removal of NCT, SD and mean of shear modulus were reduced by 6.6 and 2.0%, respectively. SD and mean of shear modulus in ECU were about 2 and 4.5 times of values in TRI, respectively. There was also an interaction of NCT and muscle on shear modulus for both SD (P < 0.01) and mean (P < 0.05). In ECU, all subjects had reductions in SD, and all but one subject had reductions in mean when NCT was removed. The effects of NCT were more variable in TRI compared with ECU. With the removal of NCT, SD and mean were reduced by 7.3% (P < 0.01) and 2.3% (P < 0.01) in ECU, respectively, but the smaller and more variable reductions in TRI (by 4.1% in SD and 0.8% in mean) did not reach statistical significance (P > 0.05).
Fig. 5

Standard deviation (SD, top row) and mean (bottom row) of shear modulus with and without non-contractile tissues (NCT) in ROI for each video in each muscle. Elastography data from ECU, TRI, and all videos are shown in the left, middle, and right columns, respectively. **P < 0.01

3.3 Depth-related distribution of NCT removal

There was a significant interaction of depth and muscle on NCT removal (P < 0.01) (Fig. 6). In TRI, more than half of NCT removal occurred at the shallowest depth whereas only 10% or less of NCT removal occurred at other depth levels, on average. The NCT removal from the shallowest level was greater than all but the 2 deepest levels (P < 0.05) where variability was large in TRI. In ECU, NCT removal was 5–14% across depth levels, on average. NCT removal from the shallowest depth was not significantly different from any other depth level except for the next shallowest depth level (P < 0.05) in ECU. At the shallowest level, NCT removal in TRI was more than four times compared with ECU (P < 0.05). Conversely, NCT removal from TRI was less at the deepest 7 levels compared with ECU (P < 0.05).
Fig. 6

Distribution of NCT data removal in ROI by depth. Each graph shows the percentage of all removed pixels that were removed from a given depth in ROI

3.4 Total area of NCT

The total area of NCT was expressed as a percentage area relative to the whole area in the ROI (NCT area) of the B-mode images in each muscle. On average, NCT area in TRI (13.5 ± 9.78%) was less than one-third of that in ECU (44.87 ± 12.2%, P < 0.001). There was no significant correlation across subjects between NCT area and the change in mean or SD of shear modulus due to the removal of NCT when data for ECU and TRI were analyzed together or independently.

3.5 Linearity of elastography with loading

Muscle elasticity values for each subject, muscle, and task with and without NCT, and the Pearson coefficients for elasticity across the linear loading conditions are displayed in Table 2. Removal of NCT increased the Pearson coefficient of mean muscle elasticity in 7 of 11 subjects for both ECU and TRI (6 of 11 in TRI when taken to 3 decimal places). In subject-muscle conditions for which the linearity was already moderate (ECU in S4 and TRI in S7) or low (ECU in S8 and S11) the removal of NCT had larger but varying effects on linearity.
Table 2

Effect of NCT removal on the linearity (r) of muscle elasticity across linear loading conditions in each subject

Data

Task

Subjects

S1

S2

S3

S4

S5

S6

S7

S8

S9

S10

S11

ECUw/NCT

2

165

237

174

86

155

180

183

217

184

181

342

1

253

273

314

118

312

278

334

233

266

202

392

3

279

313

340

101

324

298

325

222

284

273

359

r

0.955

0.999

0.928

0.457

0.897

0.934

0.838

0.289

0.939

0.954

0.335

ECUw/o NCT

2

163

227

161

112

143

161

169

218

172

172

340

1

249

281

298

144

312

270

324

244

246

178

377

3

278

328

356

107

326

280

321

230

269

262

355

r

0.963

0.999

0.974

− 0.148

0.899

0.902

0.860

0.471

0.957

0.893

0.387

TRIw/NCT

2

19

12

31

8

16

15

68

17

4

6

36

1

58

113

52

7

68

19

87

53

12

40

76

3

56

113

54

33

95

65

79

58

29

73

88

r

0.842

0.866

0.909

0.862

0.982

0.900

0.576

0.917

0.976

1.000

0.951

TRIw/o NCT

2

14

10

32

8

16

15

75

17

4

5

32

1

44

95

55

8

68

19

91

52

12

40

75

3

49

103

57

35

91

65

82

58

30

72

87

r

0.924

0.903

0.911

0.864

0.976

0.900

0.444

0.927

0.977

1.000

0.951

Cases where removal of NCT increased the r value are in boldface

3.6 Rank order of tasks

Due to the complex loading configurations and the presence of multiple degrees of freedom in the activation patterns of multiple muscles for resisting against each loading configuration, the amount of activation in ECU and TRI for each configuration was expected to be variable between subjects. Accordingly, the rank order of tasks was determined based on mean shear modulus in each subject. When NCT was not removed, the mean shear modulus at the lowest and highest ranks were 59.6 ± 19.1 and 122.8 ± 32.0 kPa for ECU and 3.2 ± 1.9 and 47.6 ± 20.6 kPa for TRI, respectively. When NCT was removed, they were 56.8 ± 19.2 and 123.2 ± 29.3 kPa for ECU and 3.1 ± 2.0 and 48.2 ± 21.1 kPa for TRI, respectively. The differences in rank order of the tasks by mean shear modulus due to the inclusion/exclusion of NCT were determined within subjects in each muscle. Assuming that the rank order based on the values without NCT is correct, the number of non-matching instances (i.e. error) was counted. Out of 13 ranks, the error at each rank ranged from 1 to 6 instances for ECU and from 0 to 6 instances for TRI, respectively. In total, rank errors occurred in 51 instances for ECU and 33 instances for TRI, respectively, out of 143 possible instances across all tasks and subjects (Fig. 7). These rank errors corresponded to the error rate of 36 and 23% in ECU and TRI, respectively.
Fig. 7

Histogram of errors in determining the rank order of tasks based on mean shear modulus. For each subject, tasks are ranked from lowest to highest mean shear modulus with and without NCT. Rank order without NCT is displayed along the x-axis. The number of subjects for which the rank without NCT differs from the rank with NCT is tallied for each task ranking

4 Discussion

The current study developed image processing algorithm for automatically removing NCT from ultrasound elastography videos of contracting muscles and found substantial effects of removing NCT on the assessment of muscle shear modulus in ECU and TRI, including the changes in the rank order of mean shear modulus across various muscle loadings.

4.1 Image processing algorithm

For identifying NCT from an ultrasound image, we employed a clustering algorithm based on the intensity values of the B-mode image. In the musculoskeletal system, muscle pennation angle, thickness and volume have been estimated from ultrasound and MRI images both manually and with automatic image processing (Graf et al. 1999; Loram et al. 2006; Zhou and Zheng 2008). For image processing, these features are often captured using edge or line detection algorithms such as the Hough transform (Zhou and Zheng 2008), intensity gradients (Graf et al. 1999), and other standard methods. For identifying NCT from an ultrasound image, a method that identifies areas instead of edges and lines is more appropriate. This sort of image segmentation has been performed primarily based on intensity values (Yang et al. 2002) or, if intensity values do not differ, visual texture of the captured tissues (Kim et al. 1998). As the tissues of interest here were distinguishable by their differing intensities, we chose a clustering algorithm based on these intensity values. Clustering has been used to capture multiple tissues by minimizing the combined intra-cluster variance (kmeans), with some variations (cmeans, a probabilistic grouping)(Yang et al. 2002). For our algorithm, Otsu’s method was implemented for just two groups, NCT and contractile muscle tissues. This clustering produced an intensity threshold separating areas of high-intensity pixels (NCT) and low-intensity pixels (contractile muscle tissues). As Otsu’s method alone was not expected to provide adequate NCT detection (≥ 90% correctly identified) across subjects and muscles, optimization was used to modify the intensity threshold and elongate the clusters to reflect the band-like shape of the NCT.

By building upon established clustering methods, the currently algorithm was developed to produce the largest amount of correctly identified NCT and contractile tissue (Results: 90.01% of NCT, 81.46% contractile tissue). Optimization of the involved parameters for multistep culling and lengthening of identified tissues enhances the segmentation. Allowing the optimization to dictate the thresholds and multipliers showed that Otsu’s method itself did not produce optimal results as none of the parameter optimization iterations selected the parameter set. Further, 23% less NCT was identified without optimization. Collectively, the addition of optimization beyond basic black-white thresholding provides additional accuracy of tissue segmentation for removal of NCT and is therefore expected to optimize the calculation of muscle shear modulus.

4.2 Elastography data

The wide range of shear modulus confirms that subjects produced various levels of muscle contractions in ECU and TRI for holding a robotic handle that received different combinations of forces and torques. The range of shear modulus across loading configurations was about twofold from 60 to 123 kPa for ECU and about 15 times from 3.2 to 48 kPa for TRI. In the triceps brachii, shear modulus at rest has been reported as 3-10 kPa (Akagi et al. 2015; Lacourpaille et al. 2012), suggesting that some of the tasks in this study may have left the triceps at or close to a resting state. Though there is no report on shear modulus of the contracting triceps muscle, shear modulus up to 47 kPa in the current study is within the reported changes in the contracting elbow flexors: 5–60 kPa during contraction from 0 to 35% MVC (Bouillard et al. 2012) and at 25–200 kPa during contractions from 0 to 70% MVC (Lapole et al. 2015). While there is no report on shear modulus for resting or contracting ECU, the lowest value of 60 kPa in the current study is higher than the resting value in other muscles (10 kPa or below). It is likely that ECU was contracting across tasks as subjects were grasping the robotic handle throughout.

The appropriateness of the algorithm for NCT removal was demonstrated as a reduction in spatial variability (i.e. SD) of shear modulus within ROI across muscles. Since shear moduli of NCT are distinct from contractile tissues (Kot et al. 2012; Lacourpaille et al. 2012), the variability of shear modulus within ROI was expected to be decreased with the “outlier” NCT removal. The amount of reduction in spatial variability of shear modulus is influenced by both the amount of outliers and the deviation of outliers from the shear modulus of the contractile tissue. The significant and greater amount of SD reduction in ECU compared with TRI after NCT removal is in line with the greater amount of outliers in ECU as its total amount of NCT area was greater compared with TRI. The deviation of outliers from the shear modulus of the contractile tissue can be altered depending on the contraction level of the contractile tissues. More variable reduction in SD across subjects in TRI compared with ECU may be related to the greater range of shear modulus across loading configurations that may have allowed for greater variability of contraction level in TRI. It appears that the currently developed algorithm for automatic NCT removal is sensitive enough to significantly influence the variability of shear modulus within ROI across muscles with different levels of contractions.

The impact of NCT removal on the assessment of shear modulus was expected as a reduction of mean shear modulus, assuming most of the NCT within images are connective tissues. Despite large variability of stiffness and NCT content across subjects, a significant main effect of NCT removal on reductions in mean shear modulus was shown (Fig. 5) in support of our hypothesis. With a significant interaction, the reduction in mean shear modulus due to NCT removal was significant in ECU, though not in TRI, indicating the reducing effect of NCT removal was greater and more consistent in ECU. Variations in the change in mean shear modulus with the removal of NCT can be due to (1) the amount of intramuscular connective tissue content and (2) the amount of adipose tissue, both of which are removed as NCT based on their high-intensity pixels. The former has high shear modulus while the latter has low shear modulus. The presence of many subjects who showed an unexpected increase in mean shear modulus after the removal of NCT in TRI implies that many TRI images appear to include large areas with lower shear modulus tissues, likely adipose tissue in the superficial level, and removal of this tissue counteracts the expected reduction of mean shear modulus via removal of higher shear modulus connective tissues. In support of this implication, NCT at the most superficial level made up the majority of overall NCT removal in TRI. Further, removal of superficial NCT was greater in TRI compared with ECU. These results are in favor of the suggestion that superficial adipose tissue with low shear modulus constituted the majority of NCT removal in TRI, which would counteract and prevent the decrease in mean shear modulus. Considering these examples of differences between muscles, it is conceivable that NCT removal may have various effects on mean shear modulus depending on individual cases, including the differences in individuals, muscles, constituents of muscles, location of ROI, and muscle contraction levels.

The validity of ultrasound elastography for capturing changes in muscle force has been established in previous studies showing the linear relationships amongst elasticity, joint torques, and EMG (Bouillard et al. 2011; Yoshitake et al. 2014; Nordez and Hug 2010; Koo et al. 2013). Here, we revisit the use of ultrasound elastography, but under complex loading conditions and include a muscle from which EMG data may be unreliable due to crosstalk. While this study utilized complex loading conditions, a set of linear loading conditions (i.e. Tasks 1, 2 and 3) were included to provide a more standard assessment of results produced by the algorithm. That the majority of subjects and muscles showed a linear relationship between the linear loading conditions and their respective mean muscle elasticity values, which was improved with NCT removal, suggests that the measurements and algorithm produced valid results. Since the linearity with joint torque is higher for mean shear modulus compared with EMG amplitude (Yoshitake et al. 2014), mean shear modulus with NCT removal would provide the highest linearity with joint torque and thus muscle force.

As elastography is expected to be used as an alternative measure of muscle contraction level (Nordez and Hug 2010), it is of great importance to understand how methods of image processing affect the assessment of relative muscle shear modulus across tasks or conditions within individuals. In particular, the assessment during variable loading tasks in interacting with a robotic system would be sensitively benefited by such understandings. Removal of the NCT was expected to improve the accuracy of assessing differences in contraction level by removing the “noise” introduced by data from the non-contractile elements. The removal of this noise appears to have benefited the linearity of elasticity data with linear loading,. In seven of eleven subjects in both TRI and ECU, the Pearson coefficient was increased. In other tasks, linearity was similar or slightly diminished. In subjects and muscles, the algorithm may have induced non-linearity or revealed the true non-linearity of the loading. Further analyses of these cases will be explored in future work.

Regarding variable loading, removal of NCT in ECU altered the rank order of tasks by mean shear modulus in 36% of the tested cases within the twofold distribution across tasks. In TRI, it was altered in 23% of the tested cases within the 14-fold distribution across tasks. The lower occurrence of rank alteration in TRI compared with ECU may partially be due to the wider distribution of shear modulus across tasks in TRI. Nonetheless, the findings indicate that one out of 3–4 cases can be inappropriately ranked for mean shear modulus within individuals unless the NCT are removed. This finding is even more notable given that changes in mean shear modulus after NCT removal were small when the average values across subjects were compared. That such a large portion of tasks is ranked differently depending on NCT inclusion suggests a significant impact of NCT on the comparison of muscle contraction level across tasks or configurations within individuals. In considering the potential clinical application of elastography in the assessment of a treatment in individual patients, for example, such a substantial impact of NCT within individuals would be critical.

Removing connective tissues has implications for the selection of a measurement area within the region of interest, as the measurement area would usually be a continuous region positioned in a region of uninterrupted contractile tissues. The included Q-BOX software in the ultrasound elastography device allows for various shaped measurement areas of variable size, but uses all the values in the selected areas in its calculation. The vast majority of elastography studies are conducted on larger muscles of the proximal arm or leg, where there is ample muscle volume from which to take shear modulus measurements (Bouillard et al. 2012; Akagi et al. 2015; Yoshitake et al. 2014; Lapole et al. 2015). Although the report is on animals (cow forelimb), distal muscles appear to have greater amounts of connective tissues compared with the proximal muscles (Purslow 2002). In small muscles or where there is higher concentration of NCT in the distal muscles like ECU, available uninterrupted contractile tissues may be limited and potential measurement areas will likely be closer to bands of connective tissues. Our image processing algorithm would especially be helpful in these situations where a selection of more flexible and appropriate measurement area with NCT removal is required. These considerations are also applicable to configurations where muscles are likely to lose contractile tissues and/or gain NCT (e.g. limb immobilization, muscular dystrophy), especially for the assessment within individuals.

Identification of NCT and removal of corresponding elasticity values from the calculation of mean shear modulus was performed with the current assumption that the presence of NCT did not significantly affect measurements in the surrounding contractile tissues. The examination of this assumption is beyond the scope of the present study geared toward application and would require a systematic modeling and/or experimental study in a different research design. The focus of this work was to provide an improvement upon the conventional (manual) method of image analysis and data recording. The developed algorithm has fundamental features that may be applied beyond the scope of the current study. For shear modulus determination, the direct application of the algorithm to resting muscle is possible, in which analysis of a single frame would be sufficient. The application of the NCT identification portion would be generalizable to B-mode images from other ultrasound systems as well and will likely provide an improvement over a simple application of Otsu’s intensity based thresholding method. Efforts were made not to use absolute cutoff values for the algorithm, but to instead use multiplier and quantiles that would naturally adjust to the image quality and architecture of the muscle of interest. Only small adjustments, if any, would be needed when applied to other muscles or used on images from other systems.

While efforts were made to create a generalizable algorithm, adjustments may be required depending on the subject population being studied. Male athletes were recruited for this study due to their expected ability to stabilize their muscles under complex loading and the low adipose tissue content of their limbs. These attributes contributed to greater ultrasound image quality. The current results from this population show that the developed algorithm worked and that NCT removal influenced the comparison of mean shear modulus across motor tasks. However, subject and muscle-dependent effects of NCT removal were still present in this ostensibly homogenous population and were incorporated into the algorithm during optimization. The application of this algorithm for non-athlete populations, likely in muscles with smaller thicknesses and greater adipose tissue content, may mimic results seen in ECU in this study due to the high concentration of NCT. If image quality or contrast is lower in these and other populations (e.g. older adults), the current algorithm serves as a foundation from which to make adjustments.

5 Conclusion

As the use of human–robot interaction gains popularity for neuromotor assessment, rehabilitation, and training, advanced methods of assessing muscle activity are needed. The limitations of EMG necessitate the use of alternative methods for accurate recording of muscle activity during complex loading, such as ultrasound shear wave elastography. In this work, we develop an algorithm for automated image processing. Our algorithm with an optimized clustering and thresholding algorithm successfully removed NCT from muscle elastography videos as supported by reductions in the spatial variability of shear modulus in the region of interest in forearm and upper arm muscles. In addition, we provide insights on how the inclusion of NCT data affects comparisons of muscle activity across tasks. Removal of NCT can alter the spatial mean value of shear modulus and altered the rank order of various static motor tasks based on mean shear modulus when compared with rankings with NCT included. These results suggest that NCT can be automatically removed in the analysis of elastography, and NCT removal has a substantial impact on the comparison of mean shear modulus across motor tasks with various loading configurations.

Notes

Acknowledgements

Special thanks to Dr. Tsukasa Ogasawara, Dr. Atsutoshi Ikeda, and Kazuya Aomoto for handling all aspects of controlling the robotic manipulator and for assisting with data collection. This study was supported in part by the National Science Foundation (IIS1142438, OISE-1209539) and Nakatomi Foundation.

References

  1. Akagi, R., Tanaka, J., Shikiba, T., Takahashi, H.: Muscle hardness of the triceps brachii before and after a resistance exercise session: a shear wave ultrasound elastography study. Acta Radiol. 56(12), 1487–1493 (2015)CrossRefGoogle Scholar
  2. Beck, T.W., Housh, T.J., Johnson, G.O., Weir, J.P., Cramer, J.T., Coburn, J.W., Malek, M.H.: Mechanomyographic amplitude and mean power frequency versus torque relationships during isokinetic and isometric muscle actions of the biceps brachii. J. Electromyogr. Kinesiol. 14(5), 555–564 (2004)CrossRefGoogle Scholar
  3. Bercoff, J., Chaffai, S., Tanter, M., Sandrin, L., Catheline, S., Fink, M., Gennisson, J., Meunier, M.: In vivo breast tumor detection using transient elastography. Ultrasound Med. Biol. 29(10), 1387–1396 (2003)CrossRefGoogle Scholar
  4. Bercoff, J., Tanter, M., Fink, M.: Supersonic shear imaging: a new technique for soft tissue elasticity mapping. IEEE Trans. Ultrason. Ferroelect. req. Control 51(4), 396–409 (2004)CrossRefGoogle Scholar
  5. Bernstein, N.A.: The co-ordination and regulation of movements. (1967)Google Scholar
  6. Bouillard, K., Hug, F., Guével, A., Nordez, A.: Shear elastic modulus can be used to estimate an index of individual muscle force during a submaximal isometric fatiguing contraction. J. Appl. Physiol. 113(9), 1353–1361 (2012)CrossRefGoogle Scholar
  7. Bouillard, K., Nordez, A., Hug, F.: Estimation of individual muscle force using elastography. PLoS One 6(12), e29261 (2011)CrossRefGoogle Scholar
  8. Brosnan, T., Sun, D.-W.: Improving quality inspection of food products by computer vision—a review. J. Food Eng. 61(1), 3–16 (2004)CrossRefGoogle Scholar
  9. Clark, D.J., Ting, L.H., Zajac, F.E., Neptune, R.R., Kautz, S.A.: Merging of healthy motor modules predicts reduced locomotor performance and muscle coordination complexity post-stroke. J. Neurophysiol. 103(2), 844–857 (2010)CrossRefGoogle Scholar
  10. Crowninshield, R.D., Brand, R.A.: A physiologically based criterion of muscle force prediction in locomotion. J. Biomech. 14(11), 793–801 (1981)CrossRefGoogle Scholar
  11. Darby, J., Li, B., Costen, N., Loram, I., Hodson-Tole, E.: Estimating skeletal muscle fascicle curvature from b-mode ultrasound image sequences. IEEE Trans. Biomed. Eng. 60(7), 1935–1945 (2013)CrossRefGoogle Scholar
  12. Du, C.-J., Sun, D.-W.: Recent developments in the applications of image processing techniques for food quality evaluation. Trends Food Sci. Technol. 15(5), 230–249 (2004)CrossRefGoogle Scholar
  13. Gallagher, W., Ding, M., Ueda, J.: Relaxed individual control of skeletal muscle forces via physical human–robot interaction. Multibody Syst. Dyn. 30(1), 77–99 (2013)MathSciNetCrossRefGoogle Scholar
  14. Graf, S., Gariepy, J., Massonneau, M., Armentano, R.L., Mansour, S., Barra, J.G., Simon, A., Levenson, J.: Experimental and clinical validation of arterial diameter waveform and intimal media thickness obtained from B-mode ultrasound image processing. Ultrasound Med. Biol. 25(9), 1353–1363 (1999)CrossRefGoogle Scholar
  15. Jacob, S.W., Francone, C.A., Lossow, W.J.: Structure and function in man. In: Unknown Host Publication Title. (1978)Google Scholar
  16. Kim, N.-D., Amin, V., Wilson, D., Rouse, G., Udpa, S.: Ultrasound image texture analysis for characterizing intramuscular fat content of live beef cattle. Ultrason. Imaging 20(3), 191–205 (1998)CrossRefGoogle Scholar
  17. Klamroth-Marganska, V., Blanco, J., Campen, K., Curt, A., Dietz, V., Ettlin, T., Felder, M., Fellinghauer, B., Guidali, M., Kollmar, A.: Three-dimensional, task-specific robot therapy of the arm after stroke: a multicentre, parallel-group randomised trial. Lancet Neurol. 13(2), 159–166 (2014)CrossRefGoogle Scholar
  18. Kong, Y.-K., Hallbeck, M.S., Jung, M.-C.: Crosstalk effect on surface electromyogram of the forearm flexors during a static grip task. J. Electromyogr. Kinesiol. 20(6), 1223–1229 (2010)CrossRefGoogle Scholar
  19. Koo, T.K., Guo, J.-Y., Cohen, J.H., Parker, K.J.: Relationship between shear elastic modulus and passive muscle force: an ex vivo study. J. Biomech. 46(12), 2053–2059 (2013)CrossRefGoogle Scholar
  20. Kot, B.C.W., Zhang, Z.J., Lee, A.W.C., Leung, V.Y.F., Fu, S.N.: Elastic modulus of muscle and tendon with shear wave ultrasound elastography: variations with different technical settings. PLoS One 7(8), e44348 (2012)CrossRefGoogle Scholar
  21. Lacourpaille, L., Hug, F., Bouillard, K., Hogrel, J.-Y., Nordez, A.: Supersonic shear imaging provides a reliable measurement of resting muscle shear elastic modulus. Physiol. Meas. 33(3), N19 (2012)CrossRefGoogle Scholar
  22. Lapole, T., Tindel, J., Galy, R., Nordez, A.: Contracting biceps brachii elastic properties can be reliably characterized using supersonic shear imaging. Eur. J. Appl. Physiol. 115(3), 497–505 (2015)CrossRefGoogle Scholar
  23. Latash, M.L.: The bliss (not the problem) of motor abundance (not redundancy). Exp. Brain Res. 217(1), 1–5 (2012)CrossRefGoogle Scholar
  24. Lawrence, J.H., De Luca, C.: Myoelectric signal versus force relationship in different human muscles. J. Appl. Physiol. 54(6), 1653–1659 (1983)CrossRefGoogle Scholar
  25. Liao, W.-W., Wu, C.-Y., Hsieh, Y.-W., Lin, K.-C., Chang, W-y: Effects of robot-assisted upper limb rehabilitation on daily function and real-world arm activity in patients with chronic stroke: a randomized controlled trial. Clin. Rehabil. 26(2), 111–120 (2012)CrossRefGoogle Scholar
  26. Loram, I.D., Maganaris, C.N., Lakie, M.: Use of ultrasound to make noninvasive in vivo measurement of continuous changes in human muscle contractile length. J. Appl. Physiol. 100(4), 1311–1323 (2006)CrossRefGoogle Scholar
  27. Madeleine, P., Bajaj, P., Søgaard, K., Arendt-Nielsen, L.: Mechanomyography and electromyography force relationships during concentric, isometric and eccentric contractions. J. Electromyogr. Kinesiol. 11(2), 113–121 (2001)CrossRefGoogle Scholar
  28. Maher, R.M., Hayes, D.M., Shinohara, M.: Quantification of dry needling and posture effects on myofascial trigger points using ultrasound shear-wave elastography. Arch. Phys. Med. Rehabil. 94(11), 2146–2150 (2013)CrossRefGoogle Scholar
  29. Martini, F.H., Bartholomew, E.F.: Essentials of Anatomy and Physiology: Pearson New International Edition. Pearson Higher Ed. (2013)Google Scholar
  30. Mogk, J.P., Keir, P.J.: Crosstalk in surface electromyography of the proximal forearm during gripping tasks. J. Electromyogr. Kinesiol. 13(1), 63–71 (2003)CrossRefGoogle Scholar
  31. Nordez, A., Hug, F.: Muscle shear elastic modulus measured using supersonic shear imaging is highly related to muscle activity level. J. Appl. Physiol. 108(5), 1389–1394 (2010)CrossRefGoogle Scholar
  32. Otsu, N.: A threshold selection method from gray-level histograms. IEEE Trans Syst Man Cybernet 9(1), 62–66 (1979)MathSciNetCrossRefGoogle Scholar
  33. Peng, Q., Jones, R.C., Constantinou, C.E.: 2D Ultrasound image processing in identifying responses of urogenital structures to pelvic floor muscle activity. Ann. Biomed. Eng. 34(3), 477–493 (2006)CrossRefGoogle Scholar
  34. Purslow, P.P.: The structure and functional significance of variations in the connective tissue within muscle. Comp. Biochem. Physiol. A 133(4), 947–966 (2002)CrossRefGoogle Scholar
  35. Rodriguez, K.L., Roemmich, R.T., Cam, B., Fregly, B.J., Hass, C.J.: Persons with Parkinson’s disease exhibit decreased neuromuscular complexity during gait. Clin. Neurophysiol. 124(7), 1390–1397 (2013)CrossRefGoogle Scholar
  36. Salman, M., Sabra, K.G., Shinohara, M.: Assessment of muscle stiffness using a continuously scanning laser Doppler vibrometer. Muscle Nerve 50(1), 133–135 (2014)CrossRefGoogle Scholar
  37. Shinohara, M., Sabra, K., Gennisson, J.L., Fink, M., Tanter, M.: Real-time visualization of muscle stiffness distribution with ultrasound shear wave imaging during muscle contraction. Muscle Nerve 42(3), 438–441 (2010)CrossRefGoogle Scholar
  38. Shinohara, M., Søgaard, K.: Mechanomyography for studying force fluctuations and muscle fatigue. Exerc. Sport Sci. Rev. 34(2), 59–64 (2006)CrossRefGoogle Scholar
  39. Taniguchi, K., Shinohara, M., Nozaki, S., Katayose, M.: Acute decrease in the stiffness of resting muscle belly due to static stretching. Scand. J. Med. Sci. Sports 25(1), 32–40 (2015)CrossRefGoogle Scholar
  40. Ueda, J., Ming, D., Krishnamoorthy, V., Shinohara, M., Ogasawara, T.: Individual muscle control using an exoskeleton robot for muscle function testing. IEEE Trans. Neural Syst. Rehab. Eng. 18(4), 339–350 (2010)CrossRefGoogle Scholar
  41. Yang, M.-S., Hu, Y.-J., Lin, K.C.-R., Lin, C.C.-L.: Segmentation techniques for tissue differentiation in MRI of ophthalmology using fuzzy clustering algorithms. Magn. Reson. Imaging 20(2), 173–179 (2002)CrossRefGoogle Scholar
  42. Yoshitake, Y., Takai, Y., Kanehisa, H., Shinohara, M.: Muscle shear modulus measured with ultrasound shear-wave elastography across a wide range of contraction intensity. Muscle Nerve 50(1), 103–113 (2014)CrossRefGoogle Scholar
  43. Zhou, Y., Zheng, Y.-P.: Estimation of muscle fiber orientation in ultrasound images using revoting hough transform (RVHT). Ultrasound Med. Biol. 34(9), 1474–1481 (2008)CrossRefGoogle Scholar

Copyright information

© Springer Nature Singapore Pte Ltd. 2018

Authors and Affiliations

  1. 1.School of Biological SciencesGeorgia Institute of TechnologyAtlantaUSA
  2. 2.Department of Sports and Life SciencesNational Institute of Fitness and Sports in KanoyaKagoshimaJapan
  3. 3.Research Center for Advanced Science and TechnologyUniversity of TokyoTokyoJapan
  4. 4.George W. Woodruff School of Mechanical EngineeringGeorgia Institute of TechnologyAtlantaUSA
  5. 5.School of Human Movement and Nutrition SciencesThe University of QueenslandBrisbaneAustralia

Personalised recommendations