Introduction

Muscle architecture is a primary determinant of its function. A dramatic change in muscle fiber arrangement occurs with changes in sarcomeres [1–3]. The addition of sarcomeres in series lengthens the muscle fascicle [4], increases the shortening velocity [5], develops the range of motion [6] and supports force buffering and energy absorption during decelerative motion [7]. In practice, a longer FL is positively correlated with sport-specific performance [8–11]. Consequently, knee flexor weakness and a shorter FL have been shown to increase the risk of injury [12–15]. Therefore, FL is a primary parameter to determine the muscle architecture and the potential force output of the muscle. Since science and practice have paid much attention to the volume and cross-sectional area of a muscle, research into whole muscle fiber architecture has long been underrepresented and undervalued [16, 17].

To investigate muscle architecture, the most common non-invasive method is ultrasound imaging. Ultrasound appears to be a convenient, robust, and relatively inexpensive imaging technique [18]. However, ultrasound has several disadvantages: limited field of view, poor representation of all muscle fibers, difficult and time consuming to standardise [19]. This leads to difficulties in using ultrasound to access changes in muscle fiber arrangement over time [20, 21]. The fact that a change in muscle architecture is not uniform across all regions of the muscle [22], shows that an accurate method and a macroscopic view of a muscle’s fiber arrangement is important. Therefore, an innovative technology such as DTI may be of interest. DTI is a valid neuroscientific approach to measure the diffusion of water molecules along fibers to estimate potential cellular barriers representing architectural properties [23]. As muscles also contain a certain amount of water, an mDTI also provides a valid [24–26], reliable and robust [27, 28] calculation of muscle fiber metrics. In addition, this tool can be used to examine even more metrics that also describe the health status of the muscle [29–31]. Nevertheless, mDTI also has certain drawbacks: measurement of surrogate biomarkers, reconstruction of tracts varies and depends on algorithms and analysis methods, MR field strength, vendor and sequences [32] and is time-consuming to process data [18, 33].

In order to turn mDTI into a tool of practical significance, it is worth taking a look at the most common data processing steps. Among several processing methods commonly used in this field [34–36], the slice-by-slice segmentation of muscle seems to be the most laborious and operator-dependent step [37, 38]. Segmentation is necessary to extract a three-dimensional volume of interest (VOI) for each subject [38]. Since it is known that VOI-based tractography does not stop at the surface of the VOI [28] and that different segmentation techniques alter the tractography result for the lower limb muscles [37], the question is how different segmentation routines affect mDTI analysis for the human shoulder. Furthermore, it seems interesting to check the importance of segmentation for an accelerated mDTI approach. Therefore, this study aims to show the dependence of two different segmentation routines for segmentation-based analysis (SBA 1 and SBA 2) on the manually segmented muscle model volume (MV) and its impact on tractography. This allows the interpretation that multiple sides with different segmentation operators can produce similar results. Furthermore, in an exploratory way, we want to compare these results with a segmentation or model-free approach (MFA). For the reliability analysis, we focused on the calculation of muscle shape indices (MV, FL and FV) and conventional reported tensor metrics (FA, AD, RD and MD), which also reflect the arrangement and health status of a muscle [29, 30].

Materials and methods

This study analysed a dataset from an interventional trial consisting of fifteen subjects who underwent MRI [39]. The protocol was approved by the local ethics committee (Leipzig University, No: 362/21-ek) and is registered in the German Register for Clinical Trials (DRKS00032375). The study was conducted in accordance with the relevant guidelines and regulations. Written informed consent was obtained from all subjects. The authors had no access to information that could identify individual participants during or after data collection.

Muscle segmentation

Manual segmentation was based on common mDTI methods described previously. Segmentation was performed using Mimics Materialise (v.24.0, Leuven, Belgium). Two independent operators (SBA 1 and SBA 2) segmented each supraspinatus muscle. Segmentation was based on the recorded T1w sequence with 176 slices. To compare individual differences in segmentation, each operator generated an individual segmentation routine for the whole data set. The first segmentation step was to generate a base mask by setting a threshold on the grey values and brightness of the pixels to separate muscle-tendons from bony structures. Then, both operators split the basic muscle mask to separate the supraspinatus muscle from the other surrounding tissues and to proceed with manual segmentation and correction. While SBA 1 preferred manual segmentation, operator two (SBA 2) focused on the integrated interslice interpolation tool for multiple-slice edit (Fig 1). However, both operators used integrated Software tools such as thresholding, region growing, filling hole and interslice interpolation to standardize segmentation and reduce the total duration for segmentation. Finally, each VOI was smoothed with first-order Laplacian (10 iterations, smoothing factor 1.0) and exported as 3D triangular meshes STL file.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Workflow of methods.

Workflow displays different processing steps and each methods duration in minutes (‘). Segmentation-based analysis by operator 1 (SBA 1) included four major segmentation steps, operator 2 (SBA 2) displayed three steps. Model-free analysis (MFA) did not include a segmentation and used the entire field of view as seeding area for deterministic fiber tracking. Within MFA the red cross symbolises the manual exclusion of tracts outside of the highlighted supraspinatus muscle (blue color).

https://doi.org/10.1371/journal.pone.0286280.g001

Results

All MR scans could be used for data processing and were included in the analysis. Segmentation for SBA 1 and SBA 2 was performed by two independent operators. SBA 1, SBA 2 and MFA were applied to the human supraspinatus muscle of the human right shoulder.

The descriptive results (Table 1) show similar values with low variance for mDTI indices (CV ≤ 8.02). In contrast, FL and FV show higher variance (CV ≥ 19.45) and mean differences between subjects and methods (FL mean = 36.34–41.27 mm). Greenhouse Geyser corrected rmANOVA were used to assess differences and showed a main model effect only for FL (F_{1.135, 15.890} = 22.645; p < 0.001; η²_p = 0.618). Further post-hoc paired t-tests (Fig 2) revealed significant differences in segmentations with respect to MV (t₁₄ = 6.407; p = 0.001), but no significant differences in FL between the two SBA methods (Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Violin plots and mean differences between methods.

Violin plots show mean differences between segmentation-based analysis (SBA 1 [blue] and SBA 2 [red]) and model-free analysis (MFA, yellow).

https://doi.org/10.1371/journal.pone.0286280.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive results.

https://doi.org/10.1371/journal.pone.0286280.t001

When assessing operator dependence for fiber tracking outcome measures (Table 2) the ICC was found to be good (≥ 0.751). Agreement was highest between SBA 1 and SBA 2, ranging from ICC estimates between 0.975 to 0.997. As an explorative approach, the comparison of SBA to MFA revealed an ICC of 0.751 for FL. Ignoring the results for FL, the ICC values for MFA varied between 0.840 (FV) and 0.930 (MD).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Interrater reliability.

https://doi.org/10.1371/journal.pone.0286280.t002

Bland-Altman plots (Fig 3) display the mean and the difference of each method within the limits of agreement. The lowest limits of agreement were found for SBA 1 and SBA 2, with the exception of FV(3.78 ± 12.20). The highest limits of agreement show comparisons between SBA and MFA similar for FL (4.87 and 4.93 for SBA 2 vs MFA) but different for the metric FV (1.98 and 5.76 for SBA 2 vs MFA).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Bland-Altman plots and limits of agreement.

Bland-Altman plots for interrater-reliability for segmentation-based analysis by two different operators (SBA 1 and SBA 2). and model-free analysis (MFA). The x-coordinate shows the mean of two methods and the y-coordinate shows the value differences between each method. The middle black line shows the mean of the paired differences. The dashed line presents the limits of agreement.

https://doi.org/10.1371/journal.pone.0286280.g003

Discussion

The main aim of this study was to investigate how two different and independent operators, who developed their own segmentation routine, affect the tractography. The results show that two different operators in segmentation resulted in different MV with no influence on fiber tracking and with an excellent ICC (≥ 0.975) [47]. This is consistent with Forsting et al. [37] investigating differences between independent raters using pre-specified segmentation routines for the lower limb muscles. In an explorative comparison of SBA methods to a more convenient model-free fiber tracking (MFA) approach, muscle parameters displayed a good ICC (0.751–0.843) and excellent ICC for the mDTI indices (0.901 and 0.93) [47]. Bland-Altman plots (Fig 3) showed no major anomalies in the data and no over- or under-estimation of parameters between SBA and MFA.

Due to the anatomy of the supraspinatus muscle and its tendonous structure, differences in MV are to be expected. As the task of both operators was to segment only the muscle and to develop their own segmentation routine independently, they were first introduced to segmentation. The aim was to check if the two operators approached segmentation differently and if this influences a fiber tracking result since it allows the conclusion that multiple sides can reveal similar results without interaction between raters. As expected, the two operators’ routines differed in the number of processing steps, duration (30–45 minutes), details of manual drawing, and how they defined the muscle-tendon interface (Fig 1). In terms of daily clinical routine, this case may reflect reality, as there is no standardised segmentation routine trained and checked by only one expert [38]. Based on this assumption, ICC estimates were calculated in this study using a two-way random effects model [37, 38]. Therefore, the findings reported might be generalised to any random operator performing a muscle segmentation regardless of the operator’s experience and the level of detail in manual segmentation.

The observed differences in FL between SBA and MFA could be attributed to several reasons: Based on the findings of Bolsterlee et al. [26, 34] and Damon et al. [24, 48] explaining that the most valid tracts are found at the belly of the muscle, we removed tracts that obviously crossed the muscle boundaries in MFA. This may have removed relevant fibers in MFA that were included in SBA. Secondly, MFA used the entire DTI image as the seeding area for fiber tracking, which may result in different density and distribution of tracts [37]. However, this makes it difficult to compare muscle metrics from SBA to MFA and it may be even more interesting to see how accurately these approaches detect changes in muscle architecture over time. Furthermore, cadaveric dissections show that the mean FL for the human supraspinatus muscle can vary between 28 mm and 83 mm [40] depending on the region of interest and the number of fascicles measured [17]. This indicates that FL should be compared carefully and in respect to the recruited subjects and the applied analysis methods.

Since the physiological cross-sectional area (PCSA) is commonly used to calculate the force-generating potential of a muscle [16, 49], FV may also be of interest to calculate the force-generating potential of a muscle, since it represents the distribution of fibers within the muscle. In this study, FV showed high variance (CV ≥ 21.8) due to very different MV between subjects (68–95 mm³ in SBA 2). However, the FV was not significantly different between methods (p > 0.05) and showed a good ICC (≥ 0.840). Results for FA, MD, AD and RD are comparable to Forsting et al. [37]. Furthermore, our FA and RD values revealed that the recruited subjects did not have irregular values as shown for injured subjects [29, 30]. This is important to demonstrate as it appears that muscle tears have a significant effect on FL [12–15] and DTI indices [29, 30].

DTI is an innovative opportunity to have a detailed look at a muscle’s fiber architecture. Nevertheless, it has some inherent difficulties. To generate DTI, the diffusion coefficient of water is measured and a mathematical representation of its diffusion is generated in three dimensions. Diffusion should be measured in at least six directions to represent the diffusion anisotropy of the muscle. Fiber tracking therefore maps the connectivity between different points within a biological VOI [48]. However, mDTI acquisition varies between studies in terms of sequences and its settings [48], segmentations [37, 38] and data analysis methods [35, 36]. This leads to incomparable results between studies [32]. Furthermore, since diffusion data are acquired in a macroscopic range to measure the water diffusion and to estimate fibers, a fiber length is only an estimation of the actual fiber length [50]. Therefore, a careful plausibility check based on a sum of muscle and DTI metrics is required to allow interpretation of the reconstructed tracts as muscle fibers. While other non-invasive technologies such as conventional 2D ultrasound or micro-computed tomography do not require fiber reconstruction, they appear to be more convenient for daily clinical routine in terms of injury and disease. However, it should be noted that such approaches usually measure a small number of fascicles even in resistance training studies [51], which has never been quantitatively tested to represent the muscle architecture accurately [16]. As shown by Charles et al. [17], this may lead to large errors in the interpretation of the whole muscle architecture. Therefore, mDTI appears to be more valid when considering the three-dimensional architecture of the whole muscle. In addition, mDTI can be interpreted as an accurate method for assessing changes after interventions [52].

Comparing mDTI with ultrasound also raises the issue of test economics and feasibility in research and clinical contexts. As mentioned, mDTI involves very complex data processing, which guided the subject of this study. To be precise, usually mDTI contains the following processing steps recommended [34–36]: (1) manual segmentation of target muscle based on anatomical sequences; (2a) resampling and registration of DTI images to T1w images; (2b) quality control of DTI; (2c) definition of VOI-specific stopping criteria for fiber tracking; (2d) integration of the segmented muscle model as VOI; (2e) fiber tracking; (3a) generating a surface model (3b) importing and translating the coordinates of identified tracts onto the surface model; (3c) calculation of muscle parameters. This is a typical workflow in which mDTI appears to be less competitive than ultrasound imaging. Nevertheless, we know that mDTI and ultrasound should be compared carefully [53]. As ultrasound is widely used as a practical imaging tool to detect major irregularities within anatomical structures, mDTI seems to be more important for research tasks due to a high amount of data that can be collected [35]. However, if ultrasound imaging has a higher degree of standardisation in terms of positioning and analysis [19], it may result in a similar level of effort to mDTI. Since a few studies have shown that changes in mDTI parameters can help to detect muscle tears [29, 30], we assume that mDTI may be of clinical relevance. The fact that muscle architecture measurements with mDTI have been shown to be reliable [27], and robust [28] gives mDTI advantages over ultrasound imaging [21]. The fact that SBA 1 and SBA 2 did differ in terms of effort and MV but not in terms of fiber tracking outcome, suggests that automatic segmentation approaches [54, 55] may be sufficient for accurate calculations of FL.

Conclusion

This study shows that different segmentation routines do not influence tractography for the human shoulder supraspinatus muscle. This is in concordance to a study for the lower limb [37]. Furthermore, tractography without a segmentation step (MFA) shows good ICC estimates. Therefore, SBA or MFA may be advantageous depending on the needs. In the case of a muscle lesion [29, 30], MFA can be used as a first and quick decision basis for further treatment. On the other hand, SBA can be used to show training-induced changes and for a more precise analysis of muscle parameters. In terms of test economy, MFA may compete with ultrasound imaging methods. While SBA 1 and SBA 2 differed in time and detail, but not in tractography results, automatic segmentation approaches can be considered to maintain the accuracy of FL calculations and speed up analysis. As the human shoulder joint is known to be a challenge for any imaging technique, the results of mDTI were very promising and detailed. Therefore, mDTI can claim practical relevance. The evaluated mDTI processing methods could form the basis for further interventional studies and future methods. MFA is an acceptable and reliable mDTI analysis method that shortens segmentation to significantly reduce the post-processing pipeline to assess the muscle fiber architecture and muscle health as quickly as currently possible. Therefore, mDTI seems worth considering as a standard sequence within a musculoskeletal MR application in a clinical context. However, future studies need to clarify whether further post-processing steps can be shortened or automated for daily clinical use. Furthermore, it remains unclear whether MFA could be applied to other muscles with different fiber orientations. To our knowledge, this is the first study using mDTI in the human shoulder in this context.

Acknowledgments

We thank all the radiologists and medical scientists involved in this study. We also thank Frank Yeh for software support from DSI Studio.