Measuring Myofiber Orientations from High-frequency Ultrasound Images Using Multiscale Decompositions

Abstract

High-frequency ultrasound (HFU) has the ability to image both skeletal and cardiac muscles. The quantitative assessment of these myofiber orientations has a number of applications in both research and clinical examinations; however, difficulties arise due to the severe speckle noise contained in the HFU images. Thus, for the purpose of automatically measuring myofiber orientations from two-dimensional HFU images, we propose a two-step multiscale image decomposition method. It combines a nonlinear anisotropic diffusion filter (NLADF) and a coherence enhancing diffusion filter (CEDF) to extract myofibers. This method has been verified by ultrasound data from simulated phantoms, excised fiber phantoms, specimens of porcine hearts, and human skeletal muscles in vivo. The quantitative evaluations of both phantoms indicated that the myofiber measurements of our proposed method were more accurate than other methods. The myofiber orientations extracted from different layers of the porcine hearts matched the prediction of an established cardiac mode and demonstrated the feasibility of extracting cardiac myofiber orientations from HFU images ex vivo. Moreover, HFU also demonstrated the ability to measure myofiber orientations in vivo.

1. Introduction

Muscles in the human body perform the important functions of supporting movement, maintaining posture and circulating blood throughout the body. The imaging and quantification of their structure can aid in the diagnosis of normal and pathological muscle tissues. In particular, the quantitative assessment of fiber orientations from medical images is an important issue in both research and clinical examinations (Fleming et al., 2008; Cetingul et al., 2008; Pillen and van Alfen, 2011), because fiber orientations are crucial to determining the properties of motion in the human body such as heart beats, breathing, eating or simply walking (Shi et al., 2008).

In recent years, there has been an increase in the utilization of ultrasound imaging to assess human muscles in both static and dynamic conditions (Sengupta et al., 2006; Shi et al., 2008; Miyoshi et al., 2009; Krogh et al., 2013). As a noninvasive imaging modality, ultrasound has been widely used in different examinations and assessments (Qin et al., 2011; Akbari and Fei, 2012; Fei et al., 2012; Qin et al., 2013a). It has also utilized in the imaging of injury or abnormal skeletal muscles (Krogh et al., 2013; Mademli and Arampatzis, 2005; Kiesel et al., 2007), modeling and rebuilt skeletal muscle structure (Fukunaga et al., 1997; Hodges et al., 2003), and movement detection through the measurement of muscle states (Miyoshi et al., 2009). Besides these efforts in muscle imaging, additional work has been done to quantify myofiber orientations in ultrasound images. Traditionally, the myofiber quantifications were conducted by manually drawing lines in the images, which was time-consuming and subjective. Therefore, Zhou and Zheng proposed an automatic estimation method based on a revoting Hough transform to extract muscle fibers in musculoskeletal ultrasound images (Zhou and Zheng, 2008). In addition to the orientation, the muscle fiber movements in an ultrasound image series could also be automatically detected to analyze human movements (Miyoshi et al., 2009). Another approach, different from these direct measurements, was proposed to first enhance fiber structures by reducing the speckle noise in the ultrasound images (Dutt and Greenleaf, 1996; Yu and Acton, 2002; Yongjin and Yong- Ping, 2009). The myofiber orientations were then extracted by edge detection from the speckle denoised images. Unfortunately, it is still difficult to quantify small myofiber orientations in clinical ultrasound images, especially for cardiac imaging. The relatively lower ultrasound frequency (<20 MHz) leads to lower imaging resolution and makes it difficult to quantify cardiac myofiber orientation.

On the other hand, high-frequency ultrasound (HFU) with a frequency of higher than 20 M Hz, which can offer high-resolution and high-quality images (Filoux et al., 2011), has been applied to imaging small-scale tissue structures in different clinical applications such as ophthalmology (Bedi et al., 2006; Martius et al., 2010), dermatology and intravascular ultrasound (Foster et al., 2009; Foster et al., 2011; Moran et al., 2011). There have also been efforts to image myofibers using HFU (Bellah, 2001; Qin et al., 2013b). It has been utilized as a convenient, practical, and inexpensive method for evaluating a variety of musculoskeletal disorders in pediatric patients (Bellah, 2001). Qin et al. validated the feasibility of using HFU to image the cardiac fibers of porcine hearts in vitro (Qin et al., 2013b). However, current quantification methods have problems in measuring small myofiber orientations from these ultrasound images because HFU leads to a vast number of scatter points in the images (Crosby et al., 2009). These speckles easily disrupt the myofiber structures in HFU images. Because of their small size, previous denoising filters, such as nonlinear anisotropic diffusion filters and speckle reducing anisotropic diffusion filters (Yu and Acton, 2002), smoothed the images by considering these small myofibers as noise and then eliminating them, and by enhancing the contrast among different tissue regions in a larger scale. On the other hand, if the speckles are not processed, the direct edge detection methods will take the speckle edges for fiber boundaries, which extracts the wrong fiber orientations.

Therefore, we proposed an automatic myofiber extraction method based on multiscale image decompositions in order to accurately extract the myofiber orientations from the HFU images. The combination of multiscale decomposition and various filtering has been an active research area in image processing (Subr et al., 2009; Choudhury and Medioni, 2011; Nilufar et al., 2012) and has been applied to medical image enhancement and analysis (Galun et al., 2007; Yoo et al., 2008; Paquin et al., 2009; Wang and Fei, 2009; Tang and Tang, 2012). In contrast to other approaches, our proposed myofiber extraction method contains a two-step multiscale image decomposition: (1) nonlinear anisotropic diffusion filtering (NLADF) (Perona and Malik, 1990; Gerig et al., 1992), and (2) coherence enhancing diffusion filtering (CEDF) (Weickert, 1999; Burgeth et al., 2009). This method can decrease the speckle noise while enhancing the myofiber structures.

This paper is organized as follows: Section 2 describes the automatic myofiber orientation measurement method; Section 3 describes the validation results on fiber-structure phantoms and on myofiber ultrasound images; Discussion and Conclusions are summarized in Section 4.

2. Methods

2.1. HFU scanners and data acquisition

An imaging system (Vevo 2100), VisualSonics, Toronto, Canada) was used in this study. The ultrasound scanner was a dedicated, high-resolution system for pre-clinical research applications. Its ultra-high-frequency transducers had superior resolution and image uniformity through the entire field of view. The system had been used in small animal imaging for cardiovascular, epidermal, and ophthalmologic applications. One of its linear ultrasound probe with a central frequency of 40-MHz (Model MS550D) was used to image the cardiac fiber orientations. Another linear probe with a central frequency of 21-MHz (Model MS250) was applied to image skeletal muscles in human subjects.

First, in order to test the cardiac myofiber imaging, fresh porcine hearts were utilized in three different steps. A left ventricle (LV) free wall was cut into triangle shapes with a size of 12 mm × 10 mm × 3 mm. After the specimen was imaged by the MS550D 40-MHz probe, it was fixed and processed histologically. Second, a specimen from the LV free wall was split along the cardiac muscle direction into small bunches around 0.6–1 mm in diameter. These small bunches were settled in distilled water and then imaged by the MS550D 40-MHz probe. Finally, specimens from the left ventricle (LV) free wall, LV apex, septum wall, and right ventricle (RV) free wall were excised and imaged by the 40-MHz probe. During imaging, the probe was attached to the cut surface of the specimens and was rotated to image the fiber structures. The imaging parameters of these three steps were set as 40-MHz frequency, 14.08 mm × 10 mm imaging region, and 256 line density.

For the purpose of skeletal myofiber imaging, the MS250 21-MHz ultrasound probe was first placed on the right calf of a human subject and adjusted to image the clear gastrocnemius fibers. Then it was utilized to image the brachioradialis muscles in the left arm. The imaging parameters were set as 21-MHz frequency, 18 mm ×24 mm imaging region, and 256 line density.

2.2. Myofiber simulation

In order to quantitatively analyze the proposed method, different simulated HFU images are generated by Field II (Jensen and Svendsen, 1992), which linearly calculated the impulse response at all scatterer positions. Two different myofiber phantoms and corresponding imaging parameters are set to simulate the HFU images of 21-MHz and 40- MHz linear probes, respectively.

For the 21-MHz probe, its phantom is set as follows: the whole volume is 20mm×25mm×10mm and 30 scatterers per mm³ are randomly distributed in it. Five skeletal myofibers are set with the size of 4.0 mm in length and 0.1 mm in diameter. Three different reflection amplitudes of myofiber scatterers that are 3, 5, and 10 times higher than those of background are chosen to set different quality images. Similarly, for the 40-MHz simulation, its phantom volume is 16 mm×12 mm×6 mm and 130 scatterers per mm³ are randomly distributed in the phantom. Five cardiac myofibers are set with a size of 2.0 mm in length and 0.045 mm in diameter. Three of the same reflection amplitude options are chosen for different image quality. All fibers are set at the known positions and arranged perpendicularly to the ultrasound beam directions in the images. During simulated ultrasound image acquisition, the imaging parameters for 21-MHz images are set as: center frequency is 21 MHz, sampling rate is 100 MHz, and focus depth is 16mm. The imaging parameters for 40-MHz images are set as: center frequency is 40 MHz, sampling rate is 200 MHz, and focus depth is 7mm. After the RF data acquisition, all data are log-compressed into gray scale images in the dynamic range of 60 dB.

Moreover, contrast-to-noise ratio (CNR) in the log-compressed image is utilized to indicate the noise level in the simulated images. The definition of CNR is (Thitaikumar et al., 2007):

$$CNR=\frac{I_{mf}-I_{bg}}{{(Va{r}_{mf}+Va{r}_{bg})}^{1/2}}$$

where I_mf is the average gray scale intensity of myofibers in the image and Var_mf is its corresponding variance. I_bg is the average gray scale intensity of the background and Var_bg is its corresponding variance.

2.3. Myofiber extraction method

After data acquisition, a two-step multiscale image decomposition approach is proposed to extract the cardiac myofiber orientations from these images. Because we focus on the myofiber orientations, the tissue regions without myofibers were manually neglected. Figure 1 illustrates the whole flowchart of the fiber extraction method. The HFU images have serious speckle noise as shown in Figure 4(b), which increases the difficulty of extracting myofiber structures. Hence, it is necessary to decrease this noise before fiber orientation extraction. On the other hand, although the fiber structures are bright in the images, the size of the small myofibers is similar to that of the noise. It is not trivial to distinguish fibers from noise. Therefore, the two-step multiscale image decomposition approach is proposed to solve this problem. The difference between both decompositions is the filters that boost different components of the signals.

Figure 1.

Flowchart of the proposed myofiber extraction method.

Figure 4.

Comparison between high-frequency ultrasound image and the stained histology slide. (a) The specifically shaped tissue from the left ventricle wall of a porcine heart during ultrasound imaging. (b) The corresponding ultrasound image. (c) The corresponding histological slide. (d) The enlargement of one small region from the histological slide.

The general procedure for each multiscale decomposition step is illustrated in Figure 2. First, different scaled images F(t₁) and F(t₂) are calculated by nonlinear diffusion filters with different iteration times t₁ and t₂ (t₁ < t₂). Then, the input image is decomposed into different components D₁ and D₂ at different scales. After a threshold denoising for each component, the processed image is reconstructed by boosting the selected scales. The reconstructed image ReI is as follows:

$$\mathit{\text{ReI}}=k_1\cdot D_1+k_2\cdot D_2+(1-k_1-k_2)\cdot F(t_2),$$ (1)

where both k₁ and k₂ are the weighting parameters.

Figure 2.

Flowchart of the two-step multiscale image decompositions.

Figure 3(a) shows an original HFU image of left ventricle tissue. The whole procedure of the fiber orientation extraction method is illustrated in Figure 3. It contains three main steps: 1) multiscale decomposition with NLADF, 2) multiscale decomposition with CEDF, and 3) fiber extraction. The details are presented as follows.

Figure 3.

Extraction of cardiac fiber orientations of tissue from the left ventricular free wall. (a) The original high-frequency ultrasound images. (b-f) Processed images of multiscale decomposition of the first filter (F₁) - NLADF. (g-k) Processed images of multiscale decomposition of the second filter (F₂) - CEDF. (l) Binary image after the Otsu's thresholding processing. (m) Skeleton extraction. (n-o) Final extracted myofiber orientations laid over the original image to show the estimated fiber orientations.

A. Multiscale decomposition with NLADF

In the first multiscale decomposition step, an NLADF is utilized to decrease the speckle noise in the HFU images. The filtering method is based on the diffusion equation framework by considering anisotropic diffusion tensors. It has the advantage of decreasing noise but keeping the significant parts of the image content, typically edges, lines or other details, which are useful for the myofiber extraction. This filter is particularly suitable for the purpose of decreasing the speckle noise but keeping the main orientation information on ultrasound images. Its equation can be written as follows (Perona and Malik, 1990):

$$I{(\overline{x},t)}_t=div(c(\overline{x},t)\cdot \nabla I(\overline{x},t)),t\ge 0,$$ (2)

where I is the intensity at the 2D space location x̅, t is the diffusion time and represents iteration steps, ∇ is the gradient operator in the space domain, and c is the diffusion tensor decided by both space and time domain. c is defined as:

$$c(\overline{x},t)=\frac{1}{1+{(\frac{\Vert \nabla I(\overline{x},t)\Vert }{K})}^2},$$ (3)

where K is a constant value.

The original image as shown in Figure 3(a) is smoothed twice by the NLADF as shown in Figure 3(b, c). Both the fine scale component D₁ in Figure 3(d) and the coarse scale component D₂ in Figure 3(e) are generated separately. During this step, both components are denoised by setting the negative value as zero because the myofiber structure is generally brighter than other parts in the image. After this decomposition and denoising processes, a reconstructed image ReI₁ in Figure 3(f) is obtained with a larger weighting parameter of 0.6 to boost the fine scale component on the image.

After the first multiscale decomposition with NLADF, the speckle noise in the image is decreased and the fiber structures, as shown in Figure 3(f), are enhanced compared with the original images in Figure 3(a). However, some of these enhanced fiber structures are still disconnected because of the speckle noise during imaging, which misleads the extraction of fiber orientation. In order to overcome these disconnections in the denoised image, the second multiscale decomposition with CEDF is applied in the following step.

B. Multiscale decomposition with CEDF

CEDF, steered by a structure tensor, has been used to complete interrupted lines and to enhance fiber structures in images (Weickert, 1999). It is similar to the diffusion filter as described in Equation (2) but with a different diffusion tensor. Considering Equation (2), the diffusion tensor c is replaced by adapting the structure tensor J, whose definition is as follows (Weickert, 1999):

$$J(\nabla I))={K_{\rho }}^{*}(\nabla I\cdot \nabla I^T),$$ (4)

with K_ρ a Gaussian weighting function with sigma ρ. If we calculate the eigenvalues, µ₁, µ₂ (µ₁ > µ₂) and the corresponding eigenvectors V̅₁, V̅₂ of the structure tensor, the localized image orientations can be represented as its eigenvectors. Based on them, the diffusion function can be rewritten with the following structure tensor:

$$I_t=div(D\nabla I),$$ (5)

where the diffusion tensor D is constructed with the same eigenvectors:

$$D=[{\overline{V}}_1{\overline{V}}_2]\cdot \left[\begin{array}{cc}\hfill {\lambda }_1\hfill & \hfill 0\hfill \\ \hfill 0\hfill & \hfill {\lambda }_2\hfill \end{array}\right]\cdot {\left[{\overline{V}}_1{\overline{V}}_2\right]}^T,$$ (6)

and its eigenvalues are given as:

$${\lambda }_1=\alpha {\lambda }_2\{\begin{array}{cc}\alpha +(1-\alpha )\text{exp}(\frac{-C}{{({\mu }_1-{\mu }_2)}^2})\hfill & \hfill ,{\mu }_1\ne {\mu }_2;\\ \alpha \hfill & ,{\mu }_1={\mu }_2\hfill \end{array}$$ (7)

here C > 0 is a threshold and α ∈(0,1) is a small regularization parameter that keeps the diffusion tensor uniformly positive definite.

The reconstructed image in Figure 3(f) is filtered twice by the CEDF as shown in Figure 3(g, h). Both the fine scale component D₁ in Figure 3(i) and the coarse scale component D₂ in Figure 3(j) are generated separately. Similar to the first decomposition step, both components are denoised and the enhanced image ReI₂ in Figure 3(k) is reconstructed.

Using structure tensor and multiscale decomposition, CEDF has the specific ability to enhance fiber structures and their orientations. The comparison between the original and processed images shows the effect of the processing, seen in Figure 3(f) and 3(k).

C. Fiber extraction

After the two-step multiscale decomposition in Figures 3(a-k), the final processed results are first transformed into binary images (Figure 3(l)) based on its Otsu' threshold (Otsu, 1979). Then, their orientations are finally extracted from this binary image using a skeleton extraction method (Au et al., 2008), as shown in Figure 3(m). Figure 3(m) and its enlarged section Figure 3(o) demonstrate the final extracted cardiac myofiber orientations (red lines) overlaid on the original input image.

2.4. Evaluations

Quantitative performance assessment of the method is conducted by comparing the results with the corresponding gold standard for evaluation of classification and segmentation (Wang and Fei, 2012; Mafi et al., 2012; Qin et al., 2009; Lv et al., 2011). In this study, the known setting in a simulated image or manual identification of myofibers in a real image was used as the gold standard to evaluate the automatic extraction method. Mean absolute distance (MAD) and the Hausdorff distance (HD) are applied to quantitatively assess the performance of the proposed method. Suppose A and B are fiber orientations of automatic and manual results, respectively. They are represented by point sets: A = {a₁,a₂,..a_m} and B = {b₁,b₂,..b_n}, and their MAD and HD are defined as follows:

$$MAD(A,B)=\frac{1}{2}\{\frac{1}{m}\sum _{i-1}^{m}d(a_i,B)+\frac{1}{n}\sum _{j-1}^{n}d(b_i,A\left)\right\},$$ (8)

$$HD(A,B)=\text{max}\left\{{\text{max}}_i\right\{d(a_i,B)\},{\text{max}}_j\{d(b_j,A)\},$$ (9)

where d(a_i, B) = min_j ‖b_j – a_i‖. MAD can supply the global matching measurements and HD can supply the local similarity between automatic and manual results.

3. Results

3.1. Histological validation of cardiac myofiber orientations

Tissue from the left ventricular free wall of a porcine heart was used to verify the feasibility of imaging cardiac myofibers by HFU. Its shape was designed as a triangle as shown in Figure 4(a). The bright boundaries of the triangle can be identified in the ultrasound images in Figure 4(b) and also from the stained histology slide in Figure 4(c). From the enlargement of the histology slide in Figure 4(d), the myofibers in the tissue are clearly visible on the image. Thus, the corresponding bright lines in Figure 4(b) indicate the myofiber orientations. Furthermore, although there were deformations during the histologic processing, Figure 4(d) shows that the horizontal myofiber orientations mainly correspond to those fiber-structure orientations as shown in the enlarged region of the ultrasound image in Figure 4(b). This experiment demonstrated that although it was relatively difficult to identify each single myofiber in the cardiac tissue, the local myofiber orientations could be still identified as groups of bright fiber-structures using HFU imaging.

3.2. Estimations of the simulated phantoms

Six different simulated HFU images are shown in Figure 5, which simulate both skeletal and cardiac fiber images with different CNRs: 1.34, 1.04 and 0.77 for skeletal images; 2.20, 1.28 and 1.04 for cardiac images. To validate the performance of our proposed method (M3), we compared its results with those extracted by two other methods: a direct method by thresholding and skeleton extraction (M1), and a method of first smoothing the image by speckle reducing anisotropic diffusion filter (SRNDF) and then extracting the fiber by thresholding and skeleton extraction (M2). The estimated fibers in the six images, which were extracted by these three methods, are illustrated in Figure 5 as red lines in the images. Figures 5(b1-b6) show the results of M1, Figures 5(c1-c6) show the results of M2 and Figures 5(d1-d6) show the results of M3. It can be seen that the results of our proposed method M3 perform better than those from the other two methods. Especially for both images with lower CNRs (Figures 5(a3,a6)), the performance of our proposed method is much better than the other two methods. This conclusion is also supported by the quantitative analysis results in Table 1. Both evaluation parameters of MAD and HD demonstrate that the errors occurring in each simulated image by our method (M3) are lower than half of the other two methods.

Figure 5.

Fiber extraction from 4 simulated ultrasound images. (a1) was imaged by the 21-MHz scanner with the CNR of 1.34. (a2) was imaged by the 21-MHz scanner with the CNR of 1.04. (a3) was imaged by the 21-MHz scanner with the CNR of 0.77. (a4) was imaged by the 40-MHz scanner with the CNR of 2.20. (a5) was imaged by the 40- MHz scanner with the CNR of 1.28. (a6) was imaged by the 40-MHz scanner with the CNR of 1.04. (b1-b6) show the results of the direct method by thresholding and skeleton extraction, Figures (c1-c6) show the results of the method that first smoothes the image by speckle reducing anisotropic diffusion filter. (d1-d6) show the results of our proposed method. The red lines indicate the extracted fibers.

Table 1.

Evaluation of the extraction results from the simulated fiber images compared with the ground truth.

Phantoms	Mean Absolute Distance (µm)			Hausdorff Distance (µm)**
	**M1	**M2	**M3	M1	M2	M3
*Image 1	50	50	10	850	830	140
*Image 2	90	90	27	2070	1400	276
*Image 3	184	240	44	1544	1175	807
*Image 4	20	20	8	530	150	80
*Image 5	80	60	20	630	1030	440
*Image 6	109	243	42	715	2196	499
Mean±Std	89±56	117±99	25±15	1057±613	1130±674	374±268

^*Image 1, 2 and 3 are the simulated 21-MHz skeletal fiber images with different CNRs 1.59, 1.35 and 0.77, respectively. Image 4, 5 and 6 are the simulated 40-MHz cardiac fiber images with different CNRs 2.20, 1.28 and 1.04, respectively.

^**M1 is the method that uses thresholding and skeleton extraction. M2 is the method that first smoothes the images by speckle reducing anisotropic diffusion filter and then extracts the fiber orientation by thresholding and skeleton extraction. M3 is our proposed method.

3.3. Estimations of the excised fiber phantoms

Two high frequency ultrasound images of excised cardiac fiber bunches are shown in Figures 6(a1, b1). Each bunch contained several myofibers, and those appeared as brighter line structures in both images. These fibers were extracted by our proposed method as the red lines shown in Figures 6(a2, b2). Moreover, in order to quantitatively evaluate the extraction errors, the automatic extraction results were compared with the manual drawings performed by an expert. Figures 6(a3, b3) demonstrate that the extracted results by the proposed method M3 are closer to the manual ones, which are shown as green lines in the figures. Similar to the simulated image experiment, we also utilized the other two methods M1 and M2 to process these excised fiber images in order to validate the performance of the proposed method. Table 2 shows the quantitative results of all three extractions on both phantoms. The mean MAD and HD of our proposed method are 20 µm and 120 µm, respectively. On the contrary, the evaluation results of the other two methods are much higher, more than 200 µm for the mean MAD and more than 1000 µm for the mean HD.

Figure 6.

Comparison of myofiber orientations groups between the automatic and manual results in three myofiber tissue phantoms. (a1) and (b1) are the original ultrasound images. (a2) and (b2) are their corresponding extraction results (red lines) extracted by the proposed method. (a3) and (b3) show the myofiber orientations from the proposed method (red lines) and those from the manual results (green lines).

Table 2.

Evaluation of the extraction results of the excised cardiac fiber phantoms compared with the manual gold standard.

Phantoms	Mean Absolute Distance (µm)			Hausdorff Distance (µm)
	*M1	*M2	*M3	M1	M2	M3
Single fiber	150	210	20	1060	1740	90
Two fibers	280	280	20	1740	2840	150
Mean	220	250	20	1400	2290	120

^*M1, M2 and M3 are the same as those in Table 1.

3.4. Extracted myofiber orientations in muscular tissue images

First, the myofiber orientations of heart tissues were extracted from 2D ultrasound images. Three different layers (outer, middle, and inner) of tissue from the left ventricle free wall were imaged by the HFU imaging. The experimental scenario was illustrated in Figure 7(a). The ultrasound probe was placed on the surface near the apex of the heart, which was imaged three times from epicardium to endocardium. The orientations of different layers were extracted from these ultrasound images, as shown in Figure 7(d). From the images, the orientations of the three layers are different: the middle layer is parallel, and the other layers are perpendicular to each other. Furthermore, the intensity in the middle layer is brighter than the others because of the anisotropic character of the ultrasound images. Although the geometric structures and orientations of cardiac myofibers are complicated, simulation from DTI data and a generalized helicoid model were used to study the myofiber layers (Lombaert et al., 2012; Savadjiev et al., 2012; Vadakkumpadan et al., 2012). It has been reported that the general orientations can be divided into different layers, with the myofiber orientations continuously changing along the depth (Savadjiev et al., 2012). The model simulation was illustrated in Figure 7(c). In order to validate the results derived from ultrasound images, the orientations of the three layers in Figure 7(a) were compared with the ones simulated by the model. Figure 7(d) demonstrates that the orientations of the three different layers were similar to those of the model simulations.

Figure 7.

A comparison of different layer orientations between high-frequency ultrasound results and model simulations (Savadjiev et al., 2012). (a) Picture of the excised specimen from left ventricular free wall, which was imaged by ultrasound three times. (b) Diagram of the heart ventricles: left ventricle (LV) and right ventricle (RV) and the specimen in (a) was excised from LV free wall. (c) Myofiber orientations of the excised specimen from epicardium (EPI) to endocardium (ENDO) based on model simulations. (d) Similarities between the different layer orientations of the excised specimen extracted from ultrasound images by our proposed method and the ones from model simulations. The red arrow lines indicate the direction from EPI to ENDO of the free wall, the green arrow lines indicate the ultrasound imaging direction from apex to base of the heart, and three blue rectangles indicate the three different ultrasound imaging planes.

Figure 8 shows the myofiber orientations of four different myocardial tissue specimens from different parts of the heart ventricles. There are differences between the orientation patterns of the four tissue specimens, especially the helical orientation at the apex. These results indicated the possibility of using HFU to image different cardiac fibers of the heart ventricles and also of using our method to quantitatively analyze their cardiac fiber orientations.

Figure 8.

Cardiac myofiber orientations extracted by the proposed method from the high-frequency ultrasound images of four different specimens of different ventricle regions. (a) Free wall of right ventricle. (b) Free wall of left ventricle. (c) Ventricle septum wall. (d) Apex.

Second, myofiber fiber orientations of skeletal muscles were also extracted. Both images of the arm and leg skeletal muscles were acquired with the 21 MHz probe. The fiber orientations were extracted by the proposed two-scale image decomposition procedures. Figure 9 demonstrates the feasibility of the proposed method to extract skeletal myofiber orientations from 2D ultrasound images, where Figure 9(a1) is the muscle image of the gastrocnemius in the right human leg and Figure 9(b1) is the image of the brachioradialis in the left forearm.

Figure 9.

Myofiber orientations extracted from the 2D ultrasound of skeleton muscles. (a1) The ultrasound image of gastrocnemius muscles in the right human leg. (a2) the extracted myofiber orientations of (a1). (b1) The ultrasound image of the brachioradialis muscles in the left forearm. (b2) The extracted myofiber orientations of (b1).

4. Discussion and Conclusions

This paper proposed an automatic method to extract myofiber orientations from HFU images, which included two-step multiscale image decompositions. The results from simulated images, excised fiber phantoms, specimens of porcine hearts and human skeletal muscle images demonstrated that the proposed method could extract both cardiac and skeletal myofiber orientations from the HFU images.

HFU scanners were selected in our experiments because of their ability to image small-scale myofiber structures. For example, it has been reported that the dimensions of single cardiac myocytes were 130–140 µm in length and 18 µm in diameter (Gerdes and Capasso, 1995)) and that their grouping fiber sheets were about 40 µm in thickness, which are below the spatial resolution of the general clinical ultrasound imaging systems at low frequencies (<10 MHz). Thus, we chose the 40 MHz scanner to image the cardiac fibers due to its axis resolution of 40 µm. Considering both the imaging depth and larger diameter of skeletal fibers (100 µm), we chose a 21-MHz scanner to image the skeletal fibers.

However, the HFU also brought additional problems in the fiber orientation extractions. The most severe one was that HFU led to vast number of scatter points in ultrasound images (Crosby et al., 2009), as shown in Figure 4(b). This decreased the image quality, resulting in low CNRs, and also led to difficulties in fiber orientation extractions. Moreover, because of the small diameters of myofibers, their continuous structures could also be interrupted and misled by the speckles of other scatter points, which made it more difficult to extract fiber orientations. Therefore, aiming at these problems, our two-scale decomposition method was proposed. Figure 5 demonstrates these problems and also shows the extracted results of our method and the other two methods for the images with different frequencies and different CNRs. The evaluations in Tables 1 and 2 also quantitatively indicate that our proposed method performs better than other two methods. Furthermore, it can be seen from Figure 5 and Table 1 that the performances of the three methods become worse when the CNRs of the images decrease. But even so, the results of our proposed method are still the best in every image. That is because the first decomposition step of NLADF improves the fiber intensities among these speckle noises and because the second decomposition step of CEDF enhances the fiber orientations based on the first step, as demonstrated in Figure 3 above.

During ultrasound imaging of myofibers, one should consider not only the fiber size but also the relationship between the fiber orientations and the ultrasound beam direction. This has been well investigated in cardiac ultrasound imaging at low frequencies (McLean and Prothero, 1991; Holland et al., 2005; Crosby et al., 2009). It has been found that the fiber orientations affected the ultrasound image intensity: the fibers that are perpendicular to the ultrasound beam direction in the imaging plane were brightest and those that are parallel to the beam almost vanished. This relationship also works for the HFU imaging of myofibers. Thus, in our experiments, almost all fibers were perpendicular to the ultrasound beam direction for the purpose of imaging them clearly.

Our proposed myofiber orientation extraction method has several possible applications in both clinical and pre-clinical ultrasound imaging. First, the HFU imaging can be directly applied to skeletal muscle imaging in vivo, such as for diagnosis of muscular abnormalities in the legs and arms. Thus, it could be directly applied and may be able to provide quantitative analysis for clinical diagnosis, especially in pediatric clinics. In addition to providing the orientation of the skeletal muscles, our method can also provide their lengths, density and textures, which might be useful for clinical diagnosis. Second, our proposed method has demonstrated the feasibility of extracting cardiac fiber orientations from HFU ex vivo. Thus, it also has the potential to quickly image fiber orientations in cultured cardiac tissue patches (Radisic et al., 2004) and in surgical ventricular restorations (Cirillo, 2009) in the future. Finally, HFU imaging has been widely used in pre-clinical imaging of small animal hearts (Foster et al., 2011). Thus, our proposed method for HFU imaging can provide automatic quantifications in pre-clinical imaging applications.