Automated fiber-type-specific cross-sectional area assessment and myonuclei counting in skeletal muscle

Abstract

Skeletal muscle is an exceptionally adaptive tissue that compromises 40% of mammalian body mass. Skeletal muscle functions in locomotion, but also plays important roles in thermogenesis and metabolic homeostasis. Thus characterizing the structural and functional properties of skeletal muscle is important in many facets of biomedical research, ranging from myopathies to rehabilitation sciences to exercise interventions aimed at improving quality of life in the face of chronic disease and aging. In this paper, we focus on automated quantification of three important morphological features of muscle: 1) muscle fiber-type composition; 2) muscle fiber-type-specific cross-sectional area, and 3) myonuclear content and location. We experimentally prove that the proposed automated image analysis approaches for fiber-type-specific assessments and automated myonuclei counting are fast, accurate, and reliable.

myosin heavy chain (mhc), which forms the thick filament of the skeletal muscle contractile apparatus, exists in multiple isoforms that contribute to the functional diversity of muscle fibers (12, 24). Muscle fiber types are categorized based on their functional and metabolic characteristics, with contractile activity dictated, in part, by the ATPase activity of the different MHCs (3). Correlations between shortening velocity and MHCs have been demonstrated using various physiological and histochemical techniques (3, 10, 15, 19, 20). Mouse skeletal muscle expresses four different MHCs, generating the following pure fiber types: slow type I and fast types IIa, IIx, and IIb (22); humans lack the IIb gene. Type IIb and IIx fibers have the fastest shortening velocities, type I fibers have the slowest velocity, with IIa fibers being intermediate. Coexpression of two different MHC isoforms results in hybrid fibers (I/IIa, IIa/IIx, and IIx/IIb), considered transitional fibers (18). While muscle fiber type is largely predetermined, fiber type is also very adaptable, with fibers changing their MHC composition in response to exercise, overload (15, 20), inactivity (4, 17, 23), and aging (4), resulting in overall alterations in the speed of contraction of the muscle. Fast-to-slow and slow-to-fast fiber-type transitions occur in the following order I↔IIa↔IIx↔IIb, such that a muscle may be composed of a combination of different pure and mixed fiber types. Thus quantification of muscle fiber-type composition, as well as the size of individual fiber types, provides fundamental information that affects overall function of the muscle.

In addition to altering MHC isoform composition in response to mechanical loading, unloading, and injury (18), skeletal muscle adapts by modifying fiber (cell) size (21), and by recruiting resident muscle stem cells (satellite cells) to regenerate damaged fibers; central nucleation is characteristic of regenerating muscle fibers (11). During muscle hypertrophy, satellite cell-dependent myonuclear accretion occurs (13), and, similarly, in some modalities of muscle atrophy, myonuclear loss occurs (7). The ratio of myonuclei to cytoplasmic volume within a muscle fiber, known as the myonuclear domain (1), is believed to have physiological significance, as the myonuclear domain is both fiber type (12) and stimulus dependent (18). Due to their location around the periphery of normal muscle fibers, distinguishing myonuclei from interstitial cell nuclei at the light microscope level in muscle cross sections requires labeling of the sarcolemma in conjunction with nuclear labeling, followed by manual counting. Visually identifying myonuclei is relatively subjective and time consuming, so it is highly susceptible to both interindividual and interlaboratory variability, resulting in discrepancies within the literature, despite the use of similar animal models (2).

Studies designed to prevent muscle loss, promote muscle recovery, improve muscle function, and/or develop exercise or nutritional interventions to promote muscle health, require accurate quantification of muscle fiber size, MHC isoform composition, and myonuclear number to assess muscle adaptability and determine the effectiveness of the therapeutic interventions. We report here an automated method to rapidly and objectively quantify multiple morphological properties of muscle. It builds on our laboratory's previously reported approach for quantifying muscle fiber cross-sectional area (CSA) (14). To our knowledge, this is the first study to report a computer algorithm for automated assessment of muscle fiber-type composition, fiber-type-specific CSA, and myonuclei counting for calculating the myonuclear domain. We report the extensive testing and validation of our approach compared with manual annotations.

MATERIALS AND METHODS

Collecting and Processing of Muscle Tissues

All animal procedures were conducted in accordance with institutional guidelines for the care and use of laboratory animals, as approved by the Institutional Animal Care and Use Committee at the University of Kentucky. Adult (4–6 mo of age), male and female C57BL/6 mice were utilized in the present study. For quantification of centrally located myonuclei, plantaris muscle that was damaged during surgical removal of the gastrocnemius and soleus muscles was utilized. The damaged plantaris muscles were isolated 2 wk after surgery and serve as a model of regeneration in the present study. Euthanasia was accomplished by an intraperitoneal injection of 0.01 ml of Euthasol (pentobarbital sodium and phenytoin sodium) in conjunction with cervical dislocation. Plantaris and soleus muscles were excised, pinned to a cork block at resting length, covered with a thin layer of Tissue Tek OCT compound (Sakura Finetek, Torrance, CA), and then quickly frozen in liquid nitrogen-cooled isopentane and stored at −80°C until processing.

Immunohistochemistry

Antibodies (Abs) and reagents used were as follows: anti-dystrophin (1:50; Vector Laboratories, catalog no. VP D505), mouse IgG blocking reagent (catalog no. MKB-2213), streptavidin-AMCA (Alexa Fluor 350; catalog no. SA-5008) all from Vector Laboratories (Burlingame, CA); Texas Red-conjugated goat anti-mouse (catalog no. 610–109-121, Rockland Immunochemicals, Gilbertsville, PA); 4′,6-diamidino-2-phenylindole (DAPI; catalog no. D3571), goat anti-mouse IgG2b, Alexa Fluor 647 conjugated 2° Ab (1:250; catalog no. A21242), goat anti-mouse IgG₁, Alexa Fluor 488 conjugated 2° Ab (1:500; catalog no. A21121), goat anti-mouse IgM, biotin conjugated 2° Ab (1:150; catalog no. 626840; Invitrogen, Carlsbad, CA); and anti-MHC I (BA.D5), anti-MHC IIa (SC.71), and anti-MHC IIb (BF.F3) from Developmental Studies Hybridoma Study Bank (Iowa City, IA). Frozen muscles were sectioned (7 μm), air dried, and stored at −20°C.

For myonuclear counting, fresh-frozen muscle cross sections were immunoreacted with the dystrophin Ab, followed by Texas Red-conjugated secondary Ab to delineate the muscle fiber. Sections were postfixed in 4% paraformaldehyde and then stained with DAPI (10 nM; Invitrogen, Carlsbad, CA) to visualize nuclei. Three to five washes with phosphate-buffered saline were performed between each step.

For fiber typing, following dystrophin staining visualized using an Alexa Fluor 350-conjugated secondary Ab, cross sections were immunoreacted with Abs against MHC isoforms type I, IIa, and IIb, followed by immunoglobulin-specific secondary Abs conjugated to different fluorophores, quantified as described below. Three to five washes with phosphate-buffered saline were performed between each step followed by postfixation with methanol.

Manual Fiber Typing and Fiber-type-specific CSA

All images were captured with a black and white digital camera on a Zeiss upright fluorescent microscope (AxioImager M1) at ×20 magnification. For manual fiber-type quantification, Zeiss AxioVision Rel software (version 4.8, Oberkochen, Germany) with fluorescent channel-specific counting was used. Fibers were sequentially scored as positive/negative in the Alexa Fluor 488 (type IIa), Texas Red (type IIb), and Cy5 (type I) channels. Fibers that were scored as negative under all three channels were classified as type IIx. Fibers that were scored as positive on multiple channels were considered hybrid fibers, i.e., a fiber that was scored positive on both Alexa Fluor 488 and Texas Red channels was classified as a type IIa/IIb hybrid fiber.

For manual fiber-type-specific CSA, fibers were first classified by fiber type (see above), and then manually traced using the dystrophin border to assess fiber-type-specific CSA. The dystrophin boundary of individual muscle fibers was visualized in the Alexa Fluor 350 channel.

Manual Myonuclei Counting

Myonuclei were manually counted in images captured at ×20 magnification using AxioVision software to determine the number of myonuclei per fiber. A nucleus was identified as a myonucleus if it met one of the following criteria: 1) it was clearly located within the dystrophin boundary; 2) it was on the boundary facing inside the fiber; or 3) >50% of the area fell inside the dystrophin boundary. Nuclei located in the central portion of a fiber, at least one nuclear diameter away from the dystrophin-stained boundary, were counted as central myonuclei. Rapid, repeated manual switching back and forth between single-channel dystrophin images and merged dystrophin/DAPI images was used to determine the location of a nucleus as inside or outside of the dystrophin boundary. Following counting of myonuclei within an image, fiber number was quantified manually to express the number of myonuclei per fiber.

Automated Fiber Typing and Fiber-type-specific CSA

Digitized images of muscle cross sections were first automatically segmented, described in detail in Ref. 14, which enables quantification of the CSA of each fiber. Representative image features were then extracted as distinctive signatures to separate different fiber types. The final fiber typing was decided by a learned multiple class machine learning unit [support vector machine (SVM)]. The step-by-step process is described below.

Automated image segmentation.

The automated segmentation algorithm contains the following three steps: 1) boundary detection to delineate the muscle fiber boundaries (in this step, muscle fiber boundaries are modeled as intensity ridges. The likelihood measurements of the ridges are computed based on an eigen-decomposition of the computed Hessian matrix of the muscle image); 2) mathematical morphological operations to postprocess the detected geometric centers of the muscle; and 3) application of gradient vector flow deformable model to drive the contour toward the boundaries of the muscle fibers.

Our proposed automatic approach is accurate and represents a significant advancement in efficiency (reduced time for analysis from 25–40 min/image to 15 s/image), while accommodating common histochemical quantification obstacles, including biological (e.g., fibrosis) and technical (e.g., processing defects and poor staining quality) artifacts. This approach was validated for fiber CSA determination (14).

Feature extraction.

After automated segmentation, a set of discriminative image features were extracted to separate different types of muscle fibers based on color. A color histogram, representing the number of pixels that have a specific color intensity value for each segmented muscle fiber, was built in the color space containing red, green, and blue channels. The color histogram for the ith muscle fiber is denoted as x_i, and the ground-truth muscle fiber type is represented as L_i, where L_i ∈ {1, . . ., N} denotes different fiber types. In total we have six types: types I, IIa, IIx (unstained, no color), IIb, I/IIa, and IIa/IIb.

Learning-based automated fiber typing using one-against-all SVM.

Using the color histogram, SVM was used to identify different fiber types. SVM, first introduced by Cortes and Vapnik (6), was used for binary classification as follows.

Let X = {x₁, . . ., x_i, . . . , x_M} and Yⁿ = {y₁ⁿ, . . ., y_iⁿ, . . . y_Mⁿ} represent the feature vector (color histograms in our case) and their corresponding class labels, where x ∈ R^P, y_i ∈{−1, +1}, with p denoting the dimensionality of the feature vector. SVM solved the optimization problem:

$${\min }_{\mathit{w},b,{\epsilon }_i}\left(\frac{1}{2}{\mathit{w}}^T\mathit{w}+\gamma {\displaystyle {\sum }_{i=1}^n{\epsilon }_i}\right),\text{subject to:}{y}_i\left[h{(x_i)}^T\mathit{w}+b\right]\ge 1-{\epsilon }_i$$

where w and b define the hyperplanes in SVM, w^T represents the transpose of w, and where the training sample was mapped to a high dimension space with the function h(x).

Minimizing w^Tw is equivalent to maximizing the margins between the positive and negative training samples, and γ is used to regulate the training errors ε_i and the margins, as illustrated in Fig. 1. The red dots in Fig. 1 denote the positive training data, and the black dots denote negative training data. The purpose of SVM classifier is to find a hyperplane that can separate two classes based on maximum margin criterion. The decision boundary representing the dashed red lines and dashed blacked lines are w^Tx + b = +1 and w^Tx + b = −1, respectively. The margin between them is 2/||w|| (Fig. 1). Based on the requirement for maximum margin classifier, the linear separable SVM is learned by

$$maximize\frac{2}{||\mathbf{\text{w}}||},\hspace{0.17em}\text{subject to:}\hspace{0.17em}{\mathit{w}}^T{\mathit{x}}_i+b\ge 1(y_i=+1),\hspace{0.17em}\text{and}\hspace{0.17em}{\mathit{w}}^T{\mathit{x}}_i+b\prec 1(y_i=-1)$$

Based on simple mathematical deduction:

$$\mathbf{m}\mathbf{i}\mathbf{n}\mathbf{i}\mathbf{m}\mathbf{i}\mathbf{z}\mathbf{e}\frac{1}{2}||w||,\hspace{0.17em}\text{subject to:}\hspace{0.17em}{y}_i({\mathit{w}}^T{\mathit{x}}_i+b)\ge 1$$

To solve the optimization problem, we maximize the dual problem, since it is a simple convex quadratic programming problem. The dual problem and the decision boundaries involve mapping h(x) through an inner product, to define the inner product through a kernel function without defining the mapping. In our implementation, we use the linear kernel function k(x, x′) = x · x′.

Fig. 1.

An illustration of linear separable support vector machine (SVM) classifier. A SVM classifier is trained to find the decision boundary that maximizes the margins between the supporting vectors. See text for definition of terms.

Binary SVM classifier was extended to differentiate more than two types of muscle fibers using a one-against-all SVM. In one-against-all classifier, for each fiber type a binary SVM is trained by labeling the samples from the fiber type as positive samples and all the other classes as negative samples. A testing muscle fiber was assigned to a class having the maximum decision function output value among all the binary classifiers. A detailed comparison of different multiple-class SVM classifiers can be found in Ref. 8.

For the nth one-against-all SVM classifier C_n, the image feature vectors (color histograms) are presented as X, and the binary class labels are presented as Yⁿ, as we illustrated before. After training, N binary SVM classifiers were recorded as Cⁿ (wⁿ, bⁿ), n = 1, . . . N. Given a segmented muscle fiber with x representing its color histogram, the probability that this individual muscle fiber belongs to the nth fiber type was calculated by

$$\text{Prob}(L=n)=\frac{e^{W^{n}X+b^n}}{{\displaystyle {\sum }_{j=1}^{j=N}{e}^{w^{j}x+b^j}}}$$

The final muscle fiber type was decided by the one-against-all SVM classifier, which maximizes the probability L = argmax Prob (L = n), where n = 1, . . . N.

Automated Myonuclei Counting

Following fiber segmentation as described above and in Ref. 19, nuclear segmentation was performed in the following three steps. 1) An adaptive threshold t was selected based on the method proposed in Ref. 16 in the blue channel of the input image, and the results were treated as initial seeds. 2) Mathematical morphological operations, including erosion, dilation, and close, were followed to refine the initial seeds. For example, close operation was used to connect oversegmented seeds together. Only initial seeds with proper nuclear areas were kept as possible candidate nuclei. 3) To get the accurate nuclear boundaries, which are critical for accurate detection of the nuclear position with respect to the boundaries of muscle fibers, a deformable model (snake) was used to find the precise boundaries of each individual nucleus.

A snake is an active contour parametrically represented by v(s) = [x(s), y(s)], s ∈ [0,1] in the image domain. It is defined to move under the influence of internal and external forces to minimize the energy functional

$$E={\displaystyle {\int }_0^1\frac{1}{2}[\alpha {|v\prime (s)|}^2+\beta {|v″(s)|}^2]+E_{ext}[v(s)]\text{d}s}$$ (1)

where v′(s) and v″(s) are the first and second derivatives representing the internal energy. E_ext[v(s)] is the external energy that denotes gradients. By incorporating an extra pressure force, balloon snake (5) calculates its internal and external forces, respectively, as follows:

$${\text{F}}_{int}(v)=\alpha v″(s)-\beta v″″(\text{s})$$ (2)

$${\text{F}}_{ext}^B(v)=\gamma n(s)-\lambda \frac{\nabla E_{ext}[v(s)]}{\Vert E_{ext}[v(s)]\Vert }$$ (3)

where n(s) represents the normal vector to the snake at the specific point on v(s), γ is the corresponding parameter, and λ represents the weight for the normalized gradient. The first term in Eq. 3 serves as the pressure force to enforce the snake to inflate or deflate (γ determines the inflation or deflation), while the second term makes it converge to object boundaries. Because of the pressure force and smoothing, the initialization does not need to be close to the boundary, and it is robust to noise inside the objects (5).

Automated identification of central myonuclei.

Using the algorithm, the area of the nucleus was defined as A and area of muscle fiber as M, the following criteria were used to define interstitial cell nuclei, myonuclei, and central myonuclei:

1. If A ∩ M < 0.5A, this nucleus will not be counted as a myonucleus (Fig. 2, region A, represented with yellow contour).

2. If A ∩ M = A, and the distance between the center of the nucleus and the boundary is greater than the long axis of nucleus, it will be counted as a central myonucleus (shown in Fig. 2, region B, represented with red contour).

3. If 0.5A ≤ A ∩ M < A, this nucleus will be counted as a myonucleus (shown in Fig. 2, region C, represented with blue contour).

Fig. 2.

The illustration of the rigorous and quantitative criteria to define interstitial cell nuclei, myonuclei, and central myonuclei. The green contour denotes the muscle fiber boundary. A, nonmyofiber nucleus (the overlapped region is denoted as green, and it is <50% of the area of the nucleus). B, myonucleus (the overlapped region is >100% of the area of the nucleus). C, central myonucleus (the nucleus is completely inside the muscle fiber).

EXPERIMENTAL RESULTS

Fiber-type Identification and CSA Assessment

To systemically evaluate the proposed automated fiber-type frequency and fiber-type-specific CSA assessment algorithm, 20 digitized muscle images that had previously been manually quantified, were randomly selected for automated quantification, as illustrated in Fig. 3. Muscle cross sections, shown on the left side, were immunoreacted with a dystrophin Ab to outline the fiber boundary, followed by three different MHC Abs, specific for type I (pink), IIa (green), and IIb (red) MHC. Unstained (black) fibers express IIx MHC. The right side of Fig. 3 shows the automated segmentation results with fiber boundaries superimposed in green, and automated fiber-type identification shown in white font.

Fig. 3.

Examples of automated fiber typing results. The segmented muscle fiber boundaries, shown in green, are superimposed on the original images. The predicted fiber types are labeled on the images using white fonts.

The results of the manual and automated quantification methods for both fiber-type distribution and CSA values were directly compared (Table 1). In each row, the number of fibers of this specific type and the mean CSA of fibers for this type, determined by both manual annotation (“Man” row, Table 1) and using the automated algorithm (“Auto” row, Table 1) are given. The quantification difference (“Diff” row, Table 1) between the two methods shows that both the maximum total automated fiber-type identification difference and the average CSA quantification difference between manual and automated results are <1%.

Table 1.

Quantification results of automated fiber typing compared with manual annotations

	I		IIa		IIb		IIa/IIb		I/IIa		Negative (IIx)
ID No.	No. fibers	Mean CSA	No. fibers	Mean CSA	No. fibers	Mean CSA	No. fibers	Mean CSA	No. fibers	Mean CSA	No. fibers	Mean CSA
1
Man	3	920	15	1,793	13	3,456	2	2,669	0	0	1	3,499
Auto	3	942	15	1,806	13	3,492	2	2,673	0	0	1	3,586
Diff, %	0	2.32	0	0.74	0	1.04	0	0.15	0	NA	0	2.49
2
Man	4	898	26	1,384	9	2,385	5	2,591	0	0	2	1,429
Auto	4	908	26	1,390	9	2,407	5	2,605	0	0	2	1,443
Diff, %	0	1.03	0	0.41	0	0.92	0	0.55	0	NA	0	0.98
3
Man	9	1,001	25	1,512	9	2,805	4	2,428	4	952	1	2,047
Auto	9	1,010	25	1,523	9	2,821	4	2,443	4	953	1	2,057
Diff, %	0	0.90	0	0.69	0	0.57	0	0.60	0	0.13	0	0.49
4
Man	11	1,162	25	1,489	8	2,422	5	2,787	4	1,074	0	0
Auto	11	1,168	25	1,501	7	2,451	5	2,810	4	1,089	0	0
Diff, %	0	0.48	0	0.79	1	1.21	0	0.85	0	1.40	0	NA
5
Man	0	0	30	1,065	10	3,009	2	2,864	0	0	14	2,458
Auto	0	0	30	1,069	10	3,021	2	2,873	0	0	14	2,461
Diff, %	0	NA	0	0.39	0	0.37	0	0.33	0	NA	0	0.13
6
Man	12	1,282	43	1,257	7	2,767	6	1,804	0	0	0	0
Auto	12	1,289	42	1,284	7	2,796	6	1,806	0	0	0	0
Diff, %	0	0.59	1	2.20	0	1.03	0	0.09	0	NA	0	NA
7
Man	2	1,179	21	1,416	12	2,779	8	2,057	3	878	0	0
Auto	2	1,196	21	1,420	12	2,794	8	2,067	3	877	0	0
Diff, %	0	1.40	0	0.26	0	0.52	0	0.46	0	0.19	0	NA
8
Man	0	0	9	900	30	2,506	7	1,608	0	0	0	0
Auto	0	0	9	901	30	2,523	7	1,616	0	0	0	0
Diff, %	0	NA	0	0.17	0	0.69	0	0.48	0	NA	0	NA
9
Man	0	0	53	870	22	1,516	3	2,108	4	804	0	0
Auto	0	0	50	914	22	1,526	3	2,117	4	813	0	0
Diff, %	0	NA	3	5.09	0	0.63	0	0.43	0	1.09	0	NA
10
Man	0	0	35	1,274	16	2,668	2	1,984	0	0	0	0
Auto	0	0	34	1,286	16	2,680	2	2,003	0	0	0	0
Diff, %	0	NA	1	0.93	0	0.45	0	0.98	0	NA	0	NA
11
Man	0	0	6	864	35	2,176	0	0	0	0	5	1,495
Auto	0	0	6	867	34	2,164	0	0	0	0	5	1,496
Diff, %	0	NA	0	0.35	1	0.58	0	NA	0	NA	0	0.03
12
Man	8	1,533	29	1,612	5	3,115	8	2,633	0	0	0	0
Auto	8	1,545	29	1,634	5	3,144	8	2,675	0	0	0	0
Diff, %	0	0.74	0	1.38	0	0.94	0	1.57	0	NA	0	NA
13
Man	14	1,604	25	1,591	1	3,521	10	2,277	2	917	0	0
Auto	14	1,616	25	1,618	1	3,579	10	2,308	2	928	0	0
Diff, %	0	0.76	0	1.64	0	1.65	0	1.37	0	1.25	0	NA
14
Man	12	1,628	29	1,654	2	3,412	9	2,035	0	0	0	0
Auto	12	1,639	28	1,660	2	3,456	9	2,061	0	0	0	0
Diff, %	0	0.72	1	0.34	0	1.29	0	1.27	0	NA	0	NA
15
Man	6	1,735	30	1,383	4	3,244	10	2,437	0	0	0	0
Auto	6	1,746	30	1,399	4	3,281	10	2,455	0	0	0	0
Diff, %	0	0.63	0	1.15	0	1.14	0	0.72	0	NA	0	NA
16
Man	6	1,778	30	1,642	5	3,358	4	2,477	0	0	0	0
Auto	6	1,793	30	1,663	5	3,386	4	2,498	0	0	0	0
Diff, %	0	0.84	0	1.26	0	0.85	0	0.88	0	NA	0	NA
17
Man	9	1,618	34	1,512	3	2,754	7	2,639	0	0	0	0
Auto	8	1,583	33	1,531	3	2,789	7	2,674	0	0	0	0
Diff, %	1	2.11	1	1.28	0	1.27	0	1.34	0	NA	0	NA
18
Man	14	1,534	26	1,628	0	0	10	2,381	1	904	0	0
Auto	14	1,546	25	1,672	0	0	10	2,409	1	918	0	0
Diff, %	0	0.77	1	2.72	0	NA	0	1.19	0	1.55	0	NA
19
Man	11	1,634	26	1,624	3	3,209	8	2,501	0	0	0	0
Auto	11	1,647	26	1,642	3	3,244	8	2,525	0	0	0	0
Diff, %	0	0.78	0	1.08	0	1.09	0	0.95	0	NA	0	NA
20
Man	1	528	51	849	13	2,037	16	1,471	0	0	0	0
Auto	1	522	51	855	13	2,049	16	1,484	0	0	0	0
Diff, %	0	1.14	0	0.78	0	0.59	0	0.91	0	NA	0	NA
Average Diff, %	0.04	1.09	0.35	0.97	0.12	0.84	0	0.85	0	0.93	0.04	0.62

Auto, automated fiber typing; Man, manual fiber typing; Diff, average quantification difference in percent between the two methods; NA, not applicable. In each row, no. of fibers of this specific type is shown. Mean cross-sectional area (CSA) is that of fibers for this type.

After all of the fiber types were identified by the automated method, the overall fiber-type composition of the muscle was computed and represented as a pie chart (Fig. 4). This analysis showed that the mouse plantaris muscle is composed of >50% pure MHC type IIa fibers, with an additional 18% of fibers expressing IIa MHC being hybrid fibers, coexpressing either type IIb or type I MHC.

Fig. 4.

A pie graph showing the distribution of different fiber types in the digitized images.

The algorithm was also used to automatically quantify fiber-type-specific CSA distribution. The size distribution of each fiber type was approximately fitted using kernel density estimation, and the results are shown in Fig. 5. This analysis showed that type IIa muscle fibers were smaller than type I fibers, with hybrid type I/IIa fibers being of intermediate size. Fibers expressing type IIb MHC, whether pure or hybrid, exhibited large size variations, since there are no distinctive peaks in their corresponding CSA distribution curves (dark and light blue curves).

Fig. 5.

The distribution of cross-sectional areas (CSAs) with respect to different fiber types. The curves are generated using kernel density estimation. The x-axis represents the CSA, and the y-axis represents the percentage of the muscle fibers with a specific CSA value. Different colors are utilized to present different types of muscle fiber CSA distribution curves.

Myonuclei Counting

Automated myonuclei counting was performed on images of both normal (Fig. 6) and regenerating (Fig. 7) mouse muscles. In both Figs. 6 and 7, the original images are presented in A. The automated fiber segmentation results are shown in B; the boundary of each muscle fiber is delineated with green contours. After nuclear segmentation, the relative myonuclear positions were calculated based on their relationship to the segmented muscle fiber boundaries (see Fig. 2). Solid yellow regions are utilized to denote myonuclei, whereas solid white regions denote nuclei on/outside the fiber boundaries in Figs. 6C and 7C. For better illustrative purposes, we show higher magnification images (Figs. 6D and 7D) of small portions of the total regions counted in B and C in Figs. 6 and 7, to clearly distinguish myonuclei and central myonuclei (yellow contours) from other nuclei (white contours).

Fig. 6.

The illustration of our proposed automated myonuclei counting method. A: the original digitized image. B: automated segmentation of each individual muscle fiber; delineated boundaries of each muscle fiber are labeled with green contours. C: automated segmentation results of each individual nucleus. Myonuclei are represented with solid yellow, while other nuclei are labeled with solid white. D: some high magnification images that correspond to the dotted rectangle regions in B and C for better illustration purposes. Myonuclei are labeled with yellow contours, other nuclei are labeled with white contours, and muscle fiber boundaries are labeled with green contours.

Fig. 7.

The illustration of our proposed automated central nuclei detection method using regenerating mouse plantaris images. A: original digitized image. B: automated segmentation of each individual muscle fibers; boundaries of each muscle fiber are labeled with green contours. C: automated segmentation results of each individual nucleus. Central nuclei are represented with solid yellow, while other nuclei are labeled with solid white. D: some high-magnification images that correspond to the dotted rectangle regions in B and C for better illustration purposes. Central nuclei are labeled with yellow contours, other nuclei are labeled with white contours, and muscle fiber boundaries are labeled with green contours.

The automated and manual quantification approaches were first evaluated on 10 randomly selected digitized images of DAPI/dystrophin-stained normal adult mouse muscle cross sections (Table 2). Counting myonuclei is an extremely time-consuming process, requiring, on average, 30 min per image of several hundred fibers. The automated method required only 10–20 s to obtain the results shown in Table 2. In Table 2, the number of myonuclei per fiber, which is calculated using the total number of the myonuclei inside the muscle fiber divided by the total number of muscle fibers, is given. The average difference between automated and manual results is 8.61%.

Table 2.

Quantification results of automated myonuclei counting compared with manual annotations

	Man		Auto
ID No.	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	Diff, %
1	0.9344	57/61	0.8387	52/62	10.24
2	0.8761	99/113	0.7941	81/102	9.36
3	0.602	59/98	0.6262	62/99	−4.02
4	0.7333	66/90	0.7977	71/89	−8.78
5	0.7326	74/101	0.6734	66/98	8.08
6	1.1639	71/61	1.1864	70/59	−1.93
7	0.7368	84/114	0.8571	92/112	−11.49
8	0.6923	36/52	0.647	33/51	6.54
9	1.4042	66/47	0.9787	58/47	12.12
10	0.8939	59/66	1.014	69/68	−13.51
Average difference					8.61

Values are no. of myonuclei per fiber, calculated using the total no. of the myonuclei inside the muscle fiber divided by the total no. of muscle fibers.

We next directly compared the reproducibility of automated and manual myonuclear counts. Both the automated counting and manual counting (same experienced technician) were performed at two different times with an interval of several weeks. The detailed results are listed in Table 3. The automated algorithm always returned exactly the same counts, while the same technician's two counts differed by 7.8%, demonstrating increased reliability of the algorithm over manual annotations. The subjectivity of myonuclear counting is particularly evident when comparing counts obtained by different individuals quantifying the same images (Table 4). A second laboratory member, trained by the technician who generated the results shown in Tables 2 and 3, recounted seven randomly selected images for quantitative comparison. The average interobserver quantification difference was 27.04% (Table 4), further illustrating the value of a standardized, computerized method for myonuclear quantification.

Table 3.

Reproducibility of automated compared with manual methods for myonuclei counting

	Auto Counting 1		Auto Counting 2			Man Counting 1		Man Counting 2
ID No.	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	Diff, %	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	Diff, %
1	1.6056	114/71	1.6056	114/71	0	1.4722	106/72	1.3867	104/75	5.81
2	1.1970	79/66	1.1970	79/66	0	0.8088	55/68	0.9420	65/69	16.47
3	1.4545	32/22	1.4545	32/22	0	1.3043	30/23	1.3043	30/23	0
4	1.0323	96/93	1.0323	96/93	0	0.7917	76/96	0.8404	79/94	6.15
5	1.1250	54/48	1.1250	54/48	0	0.9583	46/48	1.0408	51/49	8.61
6	0.6892	51/74	0.6892	51/74	0	0.9067	68/75	0.8133	61/75	9.89
7	0.6941	59/85	0.6941	59/85	0	0.7262	61/84	0.8095	68/84	11.47
8	0.7407	60/81	0.7407	60/81	0	0.8434	70/83	0.8256	71/86	2.11
9	0.7444	67/90	0.7444	67/90	0	0.8511	80/94	0.7835	76/97	7.93
10	0.7831	65/83	0.7831	65/83	0	0.7857	66/84	0.7126	62/87	9.30
Average difference					0					7.8

Auto counting 1 and 2 represent two independent counts of the same image using the software; Man counting 1 and 2 denote the same individual's repeated counts of the same image several weeks apart. Values are the no of myonuclei per fiber, calculated using the total no. of the myonuclei inside the muscle fiber divided by the total no. of muscle fibers. The last row represents the average automated and manual counting difference (%) between two independent quantifications.

Table 4.

Quantification of inter-individual variability in manual myonuclei counting

	Individual 1 Counting		Individual 2 Counting
ID No.	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	Diff, %
1	0.6377	44/69	0.8525	52/61	33.68
2	0.4262	26/61	0.4375	28/64	2.65
3	0.5540	51/92	0.6596	62/94	19.06
4	0.4766	61/128	0.4032	50/124	−15.40
5	0.5325	41/77	0.8077	63/78	51.68
6	0.7333	33/45	0.7000	35/50	−4.54
7	0.6320	48/76	1.0256	80/78	62.28
Average difference					27.04

Values are no. of myonuclei per fiber, calculated using the total no. of the myonuclei inside the muscle fiber divided by the total no. of muscle fibers.

Because of the subjective nature of manual myonuclear counting, the automated algorithm was validated by counting myonuclei in cross sections from muscles known to have different myonuclear content (9). Seventeen images from either mouse soleus or plantaris muscles were quantified using the automated algorithm (Table 5). The automated quantification successfully segregated the muscles into two groups based on myonuclear abundance. Consistent with the literature (9), the soleus (images 1–8, Table 5) had a greater number of myonuclei than plantaris (images 9–17, Table 5) muscles (1.26 ± 0.24 vs. 0.73 ± 0.07 myonuclei/fiber). Values obtained by manual counting of the same images by the trained technician resulted in myonuclear counts that were not significantly different than those obtained with the automated method (1.17 ± 0.23 vs. 0.81 ± 0.11 myonuclei/fiber).

Table 5.

Identification of different mouse muscles based on automated myonuclei counting

	Auto Counting
ID No.	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	Muscle Type
1	1.1852	64/54	Soleus
2	1.6056	114/71	Soleus
3	1.1970	79/66	Soleus
4	1.5588	106/68	Soleus
5	1.4545	32/22	Soleus
6	1.0323	96/93	Soleus
7	0.9565	44/46	Soleus
8	1.1250	54/48	Soleus
9	0.6567	44/67	Plantaris
10	0.6892	51/74	Plantaris
11	0.6941	59/85	Plantaris
12	0.8333	55/66	Plantaris
13	0.8333	70/84	Plantaris
14	0.7407	60/81	Plantaris
15	0.7444	67/90	Plantaris
16	0.6111	55/90	Plantaris
17	0.7831	65/83	Plantaris

Values are the no. of myonuclei per fiber, calculated using the total no. of the myonuclei inside the muscle fiber divided by the total no. of muscle fibers.

We also tested the automated myonuclei counting algorithm on cross sections of regenerating muscles to specifically identify centrally nucleated fibers. Figure 7 shows a representative image ∼2 wk following injury of the plantaris muscle used for quantification of central nuclei. Manual and automated counting of central myonuclei were performed on five randomly selected images. The automated algorithm accurately identified fibers and the frequency of centrally nucleated fibers; manual and automated counts differed by only 6.15% (Table 6).

Table 6.

Quantification of central myonuclei using manual and automated methods for regenerating mouse plantaris images

	Man		Auto
ID No.	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	No. myonuclei/fiber	Total no. myonuclei inside muscle fibers/Total no. muscle fibers	Diff, %
1	0.4158	42/101	0.4466	46/103	7.40
2	0.75	72/96	0.7849	73/93	4.65
3	0.2941	55/187	0.2662	41/154	−9.49
4	0.5398	95/176	0.5367	95/177	−0.60
5	0.7733	58/75	0.84	63/75	8.62
Average difference					6.15

Values are the no. of central myonuclei per fiber, calculated using the total no. of the central myonuclei inside the muscle fiber divided by the total no. of muscle fibers.

DISCUSSION

We recently developed an algorithm that calculates average fiber CSA in images in which the fiber sarcolemma is delineated (14). The present study expands upon that algorithm to automatically quantify additional fiber properties: fiber-type and fiber-type-specific CSA.

Fiber-type and fiber CSA are commonly used to characterize relative contraction velocities, energy metabolism, and size of skeletal muscle fibers. The development of an algorithm that allows for rapid and reproducible quantification of these parameters provides for greater consistency and significant reductions in man-hours that would otherwise be necessary to collect these data. Subtle deviations in manual tracing of muscle fibers is eliminated, improving the quantitative nature of the CSA results. Furthermore, hybrid fibers are efficiently and accurately identified, allowing detailed analysis of fiber-type adaptations. The approach we describe is generally applicable, but has its limitations. Processing of tissue and fixation can be performed at the users' discretion. Similarly, different primary Abs can also be used, i.e., a primary Ab recognizing MHC isoforms from a company other than that listed in the paper would be sufficient. For determination of fiber CSA, primary Abs against any protein found in the sarcolemma or basal lamina can be used, such as dystrophin or laminin. Fluorophore-specific secondary Abs must match the emission spectra of those listed in the paper for the accurate determination of fiber CSA and fiber type. Deviations in the emission spectra of different secondary Abs will negatively impact the ability of the algorithm to correctly identify fiber type and CSA.

Myonuclei were also quantified following fiber segmentation. DAPI staining combined with dystrophin immunhistochemistry to visualize the sarcolemma generates images that currently must be counted manually to quantify myonuclear content. Counting is not only time consuming, but also extremely subjective and prone to large interindividual discrepancies (2). The blue fluorescence of DAPI intercalated into the DNA radiates from the nucleus, sometimes overlapping the dystrophin-stained sarcolemma, so relatively high magnification images must be analyzed and rules established to classify a nucleus as a myonucleus compared with an interstitial cell nucleus for manual counting. In our laboratory, we define a myonucleus as one in which >50% of the DAPI-stained structure is contained within the fiber, a rule also incorporated into the algorithm. Even within our laboratory, two individuals, one trained by the other, differ by >25% in their myonuclear counts of the same images (Table 4). Differences between laboratories will almost certainly be even larger. Moreover, the subjective nature of the counting dictates that a single person must count all images from a given experiment to evaluate changes in response to experimental perturbation. Even a single individual exhibits inconsistency when asked to recount the same images after a period of time (Table 3), thereby introducing variability and making interpretation of results more difficult. The development of an automated algorithm serves to reduce both intra- and interobserver variability in myonuclear assessment. This feature will grant a more objective analysis of myonuclear changes that can occur in muscle and also allow for easier comparisons of results between laboratories as differences in subjective observer assessments are greatly reduced. For development of the automated approach for myonuclear counting that includes machine learning described here, we used the manual counts from the most experienced member of the laboratory as “ground truth”. Following numerous iterations, the algorithm differed by only 8.61% from the manual counts (Table 2). Considering the intraindividual variability is 7.8% and the interindividual variability is 27.04%, the automated algorithm indeed provides accurate counting with great reproducibility. To validate the algorithm, soleus and plantaris muscle cross sections were analyzed, and the program accurately classified the muscle based on myonuclear counts. Thus the automated algorithm provides consistent and rapid quantification of myonuclear number.

An additional feature of the automated myonuclei counting algorithm is that it can specifically identify centrally nucleated fibers, thereby quantifying muscle regeneration. Less subjectivity exists in the identification of central compared with total myonuclei due to their location, which may account for more consistent results obtained between manual and automated counts (∼6% difference) for centrally located myonuclei compared with total myonuclei. Thus, once regeneration is underway, the algorithm accommodates altered features of regenerating muscle, including increased extracellular space, increased cell infiltration, significant fiber-size heterogeneity, accurate identification of fibers, and the frequency of centrally nucleated fibers.

Taken together, the algorithm enables objective and reproducible quantification of muscle fiber properties that will transform days of labor-intensive work into minutes with excellent accuracy and reproducibility. We are continuing to expand the algorithm to enable quantification of additional muscle fiber properties, such as shape. Currently, we are well on our way to developing a user-friendly interface so that this software can be used widely by the muscle research community.

GRANTS

This research was supported by National Institutes of Health (NIH) grants AG-34453 and AR-60701 to C. A. Peterson; the Jeane B. Kempner Postdoctoral Scholar Award to C. S. Fry; the Ellison Medical Foundation/American Federation of Aging Research Fellowship EPD 12102 to J. D. Lee; and the NIH National Center for Advancing Translational Sciences, through Grant UL1TR000117.

DISCLAIMER

The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the author(s).

AUTHOR CONTRIBUTIONS

Author contributions: F.L., C.S.F., J.R.J., C.A.P., and L.Y. conception and design of research; F.L., C.S.F., J.M., J.R.J., J.D.L., C.A.P., and L.Y. performed experiments; F.L., C.S.F., J.M., J.R.J., J.D.L., C.A.P., and L.Y. analyzed data; F.L., C.S.F., J.M., J.R.J., J.D.L., C.A.P., and L.Y. interpreted results of experiments; F.L., C.S.F., J.R.J., C.A.P., and L.Y. prepared figures; F.L., C.S.F., J.R.J., C.A.P., and L.Y. drafted manuscript; F.L., C.S.F., J.R.J., J.D.L., C.A.P., and L.Y. edited and revised manuscript; F.L., C.S.F., J.M., J.R.J., J.D.L., C.A.P., and L.Y. approved final version of manuscript.