An automatic fascicle tracking algorithm quantifying gastrocnemius architecture during maximal effort contractions

Introduction

Muscle shortening dynamics during contractions in the plantar flexors govern locomotor function in athletes, the elderly, and many patient populations (Kumagai et al., 2000; Abe, Kumagai & Brechue, 2000; Suzuki, Bean & Fielding, 2001; Mulier et al., 2003; Randhawa & Wakeling, 2013). The plantar flexors, despite their relatively small size compared to the muscles of the hip and knee, play a critical role in human locomotion. These muscles behave in a variety of different ways to minimize the energy expenditure needed to complete functional activities. During the stance phase of walking, the plantar flexors act mostly isometrically to facilitate elastic energy storage and return in the Achilles tendon (Fukunaga et al., 2001). However, running requires increased rates of shortening to do the positive work necessary to accelerate the body (Lichtwark, Bougoulias & Wilson, 2007). While computational models provide critical insight into muscle-tendon dynamics in response to small changes in plantar flexor structure (Nagano et al., 2007; Baxter & Hast, 2019), coupled muscle shortening and rotation is described to have complex “variable gearing” that is dependent on both load and speed (Azizi, Brainerd & Roberts, 2008) which are not present in computational models. Therefore, experimental measurements of muscle shortening dynamics are critical for both understanding movement biomechanics of different populations while also serving to improve and validate musculoskeletal models.

Structure-function relationships can be described by quantifying both muscle shortening dynamics and joint kinetics (Maganaris, 2003; Arampatzis et al., 2005). Ultrasound imaging is a popular tool for quantifying skeletal muscle structure and shortening dynamics in human subjects during functional tasks (Franchi et al., 2018). Isokinetic dynamometry provides a unique framework for measuring joint torques generated during isolated movements while controlling for load or velocity. The combination of these two measurement techniques enable researchers to study the link between muscle structure and function that underpins musculoskeletal modeling (Reeves & Narici, 2003; Zhou et al., 2012; Blazevich et al., 2009; O’Brien et al., 2010; Randhawa, Jackman & Wakeling, 2013).

Muscle shortening dynamics in pennate muscles are quantified using ultrasound by measuring changes in length and pennation of individual fascicles in a muscle belly. Unfortunately, analyzing ultrasound images acquired during functional tasks has proven to be time intensive and technically challenging. Fascicle length and pennation have traditionally been manually digitized using custom-written computer software (Maganaris, 2001; Magnusson et al., 2001; Barber, Barrett & Lichtwark, 2011), but this approach is a time intensive process when analyzing high frame rate ultrasound data. Automatic tracking routines have been developed to make these measurements using image processing algorithms (Cronin et al., 2011; Gillett, Barrett & Lichtwark, 2013; Zhou, Chan & Zheng, 2015). Cronin et al. (2011) leveraged an optical flow algorithm to quantify fascicle length and made the “Ultratrack” analysis software freely available (Farris & Lichtwark, 2016).

This tracking framework is freely available, has an intuitive interface, and reliably measures fascicle length. Ultratrack has been a major contribution to the field of functional muscle imaging as evidenced by its popularity in the biomechanics community (Hauraix et al., 2015; Clark & Franz, 2018; Hösl et al., 2018). This approach tracks the optical flow of the entire muscle rather than individual fascicles and aponeuroses. However, as noted by Gillett, Barrett & Lichtwark (2013) non-homogeneous deformations within the region of interest (ROI) violate an underlying assumption of optical flow, which may negatively impact its performance during maximal effort contractions due to the entire muscle moving relative to the transducer. This issue can potentially be avoided by tracking smaller ROIs within the image rather than the entire muscle. This approach has been used to automatically measure muscle stretch in the plantarflexors during quiet standing (Loram, Maganaris & Lakie, 2006), though to our knowledge it has not been used to track fascicle length or pennation.

The purpose of this study was to develop and validate a new fascicle tracking paradigm that directly tracks the fascicle and aponeuroses to provide measurements of fascicle length and pennation angle and make it freely available to the biomechanics community. We used this program to automatically track an individual muscle fascicle and the aponeuroses during maximal voluntary plantar flexor contractions performed on an isokinetic dynamometer throughout a range of angular velocities for five subjects. We made manual measurements of fascicle geometry to serve as a comparison for validation. We used these data to evaluate the repeatability, reproducibility, and agreement between automatic and manual measurements of fascicle length and pennation angle across three examiners across 3 days. We assessed these parameters using correlation testing and established an a priori threshold of r > 0.67 to demonstrate strong agreement between comparisons. In addition, we compared the performance of this new tracking program with Ultratrack on a subset of videos from one subject.

Materials and Methods

Supervised tracking in a simulated use case

While the unsupervised tracking approach is useful to test the repeatability and reproducibility of tracking across examiners, it is a poor analog for how automatic fascicle tracking is used in practice where users supervise the tracking. For example, Ultratrack includes functionality for the user to view videos prior to fascicle identification and make manual corrections (Farris & Lichtwark, 2016). While our program does not include any of these features, we sought to test the performance of our approach in a case study where users had the ability to re-initialize the program after watching a first attempt at tracking. As such, one examiner performed “supervised tracking” where the program was used to track all data from five subjects (75 videos) with the ability to re-initialize the tracking if poor tracking was observed. Poor tracking was defined as an obvious difference between the tracked fascicle and the fascicles in the ultrasound images upon review (Fig. 1E). Aeles et al. (2017) showed that watching ultrasound videos of fascicle shortening increased the reliability of manual fascicle identification relative to static images. During supervision, the examiner watched the tracked trial following program initialization, which provided information regarding fascicle geometry that was not apparent in the static first frame as well as allowed the examiner to evaluate the quality of the tracking. Upon observing a poorly tracked trial, the examiner rejected the tracking, re-identified the aponeuroses, and identified the same or different candidate fascicle based on watching the video (Fig. 1).

Agreement between automatic and manual tracking using the supervised approach was evaluated using CMCs and RMSEs. The Bland–Altman method of differences was used to evaluate the agreement between the two approaches (Bland & Altman, 1986). Bland–Altman analysis does not have a defined threshold for statistical acceptance and instead relies on establishing an acceptable threshold a priori based on application specific requirements (Giavarina, 2015). A recent review reported coefficient of variation ranging from 0% to 9.8% for measurements of either fascicle length and pennation angle in the human gastrocnemius across a range of measurement approaches (Kwah et al., 2013). As such, we established an a priori coefficient of variation value of 10% for both fascicle length and pennation angle.

To quantify whether errors increased over time due to the accumulation of errors due to tracking, we performed an analysis of the isokinetic data (30, 120, 210, and 500 °/s) where we calculated the RMSE between all manual and automatic measurements of fascicle length and pennation angle at each of the six ankle positions. We calculated this for both unsupervised and supervised tracking approaches.

Results

Automatic and manual measurements of fascicle length and pennation angle across all isometric and isokinetic conditions were found to be repeatable with strong intra-examiner agreement (ICC > 0.74, Table 1). These measurements were also found to be reproducible with strong inter-examiner agreement across all examiners and days (ICC > 0.76, Table 1). Automatic measurement of fascicle length and pennation had larger RMSE values relative to manual measurements (Table 1). Fascicle length measurements were more reproducible than measurements of pennation angle. The automatic measurements were less reproducible than the manual measurements, however, automatic tracking reproducibility and repeatability remained strong for all comparisons.

We found that the reproducibility of automatic tracking was strong for uncorrected measurements and very strong in corrected measurements of fascicle length (Figs. 2A and 2B) and pennation angle measurements (Figs. 3A and 3B). Corrected values for fascicle measurements had a higher mean CMC value with a smaller standard deviation (CMC = 0.98 ± 0.02) relative to the uncorrected values (CMC = 0.88 ± 0.09). Automatic tracking reproducibility of pennation angle was also improved following initial bias correction (CMC = 0.92 ± 0.04) compared to uncorrected values (CMC = 0.84 ± 0.1).

Agreement between automatic and manual measurements of fascicle length was very strong in both uncorrected (CMC = 0.90 ± 0.13) and corrected (CMC = 0.93 ± 0.10) values (Table 2). Isometric conditions resulted in the higher CMC values and smaller errors relative to the isokinetic conditions (Table 2). There was greater error in the uncorrected values (RMSE = 5.86 ± 4.13 mm) relative to the corrected values (RMSE = 4.97 ± 3.65 mm) (Table 2). The RMSE was highest at 500 °/s and lowest for isometric, although isometric had the highest RMSE standard deviation.

Agreement between automatic and manual measurements of pennation angle was strong in both uncorrected (CMC = 0.83 ± 0.18) and corrected (CMC = 0.86 ± 0.15) values (Table 3). Isometric conditions resulted in the higher CMC values and smaller errors relative to most of the isokinetic conditions (Table 3). There was greater error in the uncorrected values (RMSE = 7.16° ± 4.25°) relative to the corrected values (RMSE = 6.22° ± 3.90°) (Table 3). The RMSE was lowest for isometric for measurements of pennation angle in contrast to fascicle length in which it was highest.

Supervised automatic tracking demonstrated very strong agreement between manual and automatic measurements of both fascicle length (CMC = 0.97 ± 0.04) and pennation angle (CMC = 0.94 ± 0.06) as well as a considerably smaller RMSE relative to unsupervised (Table 4). Supervision sharply reduced the incidence of tracking trials with poor and moderate agreement with over 90% of fascicle length measurements and over 80% of pennation angle measurements having very strong agreement between manual and automatic approaches (Fig. 4).

The coefficient of variation for supervised measurement of fascicle length and pennation angle was 8.6% and 11%, respectively (Fig. 5). The coefficient of variation for pennation angle fell slightly outside of our a priori threshold of 10% while fascicle length met this criterion. Despite the strong agreement between the automatic and manual tracking approaches, the unsupervised automatic tracking under-approximated fascicle lengths by 5.6% and over-approximated pennation angles by 10.1% across the entire range of motion. Supervision of automatic tracking decreased these errors to less than 1.1% for under-approximation of fascicle length and 5% over-approximation for pennation angle. The values are mean signed errors.

With regards to quantifying the accumulation of error over the course of a video, we observed that RMSE in fascicle length decreased over time for both unsupervised and supervised tracking. In contrast, RMSE in pennation angle increased over time for unsupervised and supervised tracking. In both cases, supervision approximately halved the RMSE error, especially in late frames (Table 5).

Our program produced smaller tracking errors compared to Ultratrack when provided with the same starting fascicle and aponeuroses geometry for measurements of both fascicle length and pennation angle (Fig. 6). Initializing Ultratrack from a middle frame improved tracking fidelity but these errors were still greater than our approach (Fig. 7). Initializing in the middle frame reduced errors in Ultratrack in all cases outside of isometric max (Table 6).

Discussion

The purpose of this study was to establish the reproducibility, repeatability, and agreement of a novel automatic fascicle tracking algorithm that directly tracks the fascicle and aponeuroses. This algorithm was tested with five subjects during maximal voluntary contractions performed on an isokinetic dynamometer in isometric and isokinetic conditions. Our results indicate that this automatic tracking approach is repeatable, reproducible, and had strong agreement with manual measurements for three different examiners across three different days. We showed that supervision of the automatic tracking and reinitializing the program as needed provided accurate measurements of both fascicle length and pennation angle in one tracking session. Also, we showed that this tracking approach exhibited a smaller RMSE than the popular Ultratrack software when tracking the same fascicle. Our tool also provides the user with pennation angle measurements. The reliability of the automatic tracking coupled with its speed relative to manual tracking makes this approach an attractive method for studying muscle geometry during maximal effort contractions.

The agreement between automatic and manual fascicle length measurements compared well with the literature. Our repeatability measurements for manual measurements of fascicle length and pennation angle fall within previously reported values (0.87–0.99 for fascicle length and 0.8 for pennation) for measuring muscle architecture parameters in identical images (Kwah et al., 2013). The reproducibility of automatic measurements of fascicle length (0.86) and pennation angle (0.76) across all examiners and days fell just short of these values, but remained classified as strong. Previous groups have reported mean uncorrected CMC values for fascicle length measurements of 0.88 and 0.9 which compares well with our value of 0.89 (Cronin et al., 2011; Gillett, Barrett & Lichtwark, 2013). We observed that measurements of pennation angle had less agreement than measurements of fascicle length which has also been observed by previous groups (Aggeloussis et al., 2010; Kwah et al., 2013). The repeatability of automatic fascicle length measurements for uncorrected and corrected values was 0.88 ± 0.09 and 0.98 ± 0.02, respectively. This compares well with previously reported values for Ultratrack, 0.88 ± 0.08 and 0.98 ± 0.02 for uncorrected and corrected values, respectively, (Cronin et al., 2011). While there are fewer examples of automatic pennation angle measurement in the medial gastrocnemius, one group (Zhou, Chan & Zheng, 2015) reported an average correlation of r = 0.935 which compares well with our value of r = 0.94 for supervised pennation measurement. While the coefficient of variation for pennation measurements fell slightly outside of our a priori value of 10%, an 11% coefficient of variation still provides researchers with a useful tool for the automatic tracking of fascicles.

We found that the automatic tracking was reproducible across days and examiners as demonstrated by strong agreement (ICC values > 0.76). While this value was lower than the reproducibility of our manual measurements (>0.9), it represents the worst case scenario as the program was unsupervised. While ICCs were not calculated for the supervised algorithm because only one examiner performed this testing, the high correlation between manual and automatic measurements (CMC > 0.94) and reduced RMSE provides evidence supervising automatic tracking will enhance reproducibility. Automatic tracking is significantly faster than manual tracking. To initialize the program, a user only needs to draw three lines in the first frame which takes less than 15 s. To process all the frames in a 150 frame video, it takes approximately 15 s. Manual tracking, even when only digitizing six frames, took approximately 1 min. As such, automatic tracking provides measurements across the entire video in a much shorter time.

The automatic measurements were less reproducible than manual measurements in contrast to a previous study (Miyoshi et al., 2009). Automatic measurements were observed to be affected by the initial fascicle that was identified in the first frame. At times, the selected fascicle would move over a vein or another stationary structure, or the fascicle would appear to move out of the plane of the ultrasound. This would cause the tracked fascicle to “lag” and at times track the fascicle on either side of the initial fascicle. During unsupervised tracking, the first tracking attempt was accepted regardless of tracking quality, which we believe contributed to the lower agreement between manual and automatic tracking (CMC > 0.83). Supervision removes this issue because it allows the user to simply identify a different fascicle that does not have this issue (Fig. 1). The efficacy of this is demonstrated by the very strong agreement between manual and automatic measurements (CMC > 0.94) and low RMES for fascicle length (RMSE = 3.31 ± 1.53 mm) and pennation angle RMSE = 4.12° ± 1.98°). The improvement of tracking due to supervision is supported by the work of Aeles et al. (2017) where they observed that providing examiners with the ability to watch the video prior to fascicle identification improved the reliability of fascicle identification.

Similar to other studies, we found that automatic measurements of pennation angle were less reproducible than fascicle length (Zhou, Chan & Zheng, 2015) and also had higher RMSE errors. At least for the present study, we believe that this increased error in pennation is due to trigonometry. When the fascicle is contracted, small variations in fascicle length results in large variations in pennation angle (Figs. 1E and 1F). It is apparent that tracking is poor and the measurements of fascicle length (57.8 mm) and pennation angle (32°) are incorrect (Fig. 1E). Reinitializing the program and identifying a different fascicle enables better tracking (Fig. 1F). The measurements of fascicle length (50.6 mm) and pennation angle (49°) are significantly more accurate. The incorrect track results in a 12.5% over-estimation of fascicle length and 53.1% under-estimation of pennation angle relative to the well-tracked video. Aeles et al demonstrated that tracking a fascicle backward in time from contracted to relaxed state did not improve the accuracy of the fascicle length measurement relative to traditional forward tracking. This supports our results that fascicle length measurements did not degrade as the video progressed (Table 5). We observed that error in pennation angle measurements did increase over time, however. Therefore, we suggest that future studies investigating the role of tracking direction on pennation accuracy are warranted.

Automatic tracking performance was dependent on ultrasound acquisition rate and quality. During pilot testing, we acquired images at nearly half the frame-rate (30 frames/s) of our reported data (60 frames/s). We found that tracking was poor, possibly due to the large displacement of tracked points in fast contraction velocities. Based on these findings, we suggest that ultrasound images should be acquired as quickly as possible while still maintaining good image quality. Recent advances in high frame rate ultrasonography have increased the possible capture rates to 500–2,000 frames/s to study very fast ankle rotations (Hauraix et al., 2015), although the trade-off between image quality and frame rate is based on specific hardware specifications. Additionally, we found that inter-subject muscle variability affected tracking quality. Most notably, subjects with clearly identifiable fascicles in the first frame that remained visible throughout the entire contraction tracked better than trials that became less clear later in the contraction (Fig. 1). Tracking quality was partly dependent on defining the fascicle in the first frame. A potential improvement to this approach would be the development of a process to automatically identify the fascicle in the first frame to reduce variability. One possible approach demonstrated by Zhou, Chan & Zheng (2015) used images transforms to automatically identify line-like structures of the fascicle. This approach could be used to automatically establish a ROI around each of these structures in the first image.

This study was affected by several limitations. Similar to other automatic tracking approaches, we approximated muscle fascicles and aponeuroses as straight lines, which may lead to measurement errors where fascicle length is under-approximated due to fascicle curvature (Cronin et al., 2011; Gillett, Barrett & Lichtwark, 2013; Zhou, Chan & Zheng, 2015; Farris & Lichtwark, 2016). In our data, we did not observe very much fascicle curvature, even in fully contracted state. Fascicles in the gastrocnemius do not curve as much as fascicles in other muscles such as the vastus medialis (Gillett, Barrett & Lichtwark, 2013) or the biceps femoris (Seymore et al., 2017; Aeles et al., 2017). If significant amount of fascicle curvature is observed, we suggest that users exercise judgment as to whether this straight-line approximation is valid. We should note that two out of the three examiners who analyzed the ultrasound images were relative novices. While these examiners were trained by an experienced examiner, novices have been shown to be less reliable than expert and automatic tracking approaches (Miyoshi et al., 2009). However, our findings show that even novice examiners have very strong reproducibility when manually measuring fascicle length (ICC > 0.93) and pennation angle (ICC = 0.9).

Another limitation is that we did not manually digitize all of the frames across the entire contraction. Each trial contained between 50 and 150 frames which was effectively down sampled to six frames for manual measurements. Our rationale for the limited number of manual measurements was to prioritize the number of subjects (n = 5), rotational velocities (n = 5), number of trials (n = 3), and number of repeated measurements (n = 3) that each examiner (n = 3) analyzed. While other validation studies have had a larger number of manual measurements for each ultrasound video, these studies had a limited number of subjects (Miyoshi et al., 2009), a limited number of conditions (Zhou, Chan & Zheng, 2015), or digitized only a subset of the ultrasound videos (Cronin et al., 2011; Gillett, Barrett & Lichtwark, 2013). In addition, the six manual measurements were made at the same approximate time points across all contractions. As such, we were unable to closely identify in which time windows of a contraction errors occurred. We did observe that RMSE decreased over time for fascicle length measurements and increased for pennation angle measurements. Another limitation of this study is that we used an iterative tracking approach where errors in earlier frames can accumulate over the course of the track. We demonstrated that supervision of tracking increased the accuracy of the track, especially in later frames (Table 5). To reduce the accumulation of errors for longer videos in which multiple cycles of contractions are tracked, Ultratrack has a powerful set of tools to account for the propagation of errors such as keyframe adjustments and the ability to make manual corrections (Farris & Lichtwark, 2016). As our program is a command window initiated tracking program rather than a full suite of tools wrapped in a graphical user interface, we hope that our tracking paradigm can be incorporated into a tool with similar features in the future.

Conclusions

Accurately quantifying muscle architecture from ultrasound imaging during maximal effort tasks provides researchers and clinicians with an important tool for understanding the structure-function relationship that underpin muscular injury, disease, and recovery. This study introduced a novel fascicle and aponeuroses tracking approach that quantifies architectural parameters of individual muscle fascicles using a commercially available point tracker. We demonstrated that our proposed automatic fascicle tracking algorithm is repeatable and reproducible across a wide range of angular velocities even when used by different examiners across different days. It showed strong agreement with manual measurements, especially when used in a supervised manner. When initialized from the first frame with identical starting positions to Ultratrack, our program exhibited less error while also providing accurate measurements of pennation angle.