Fractal and gray level co-occurrence matrix texture analysis of senescent and non-senescent deciduous teeth stem cells: A pilot study

All cells in the human organism, including stem cells, have a limited lifespan and cease to divide after a certain number of divisions. The process of cellular senescence refers to incapability to progress through the cell cycle, thus leading the cell to an irreversible growth arrest. One of the most common biomarkers of cell senescence is senescence-associated β-galactosidase (SA-β-Gal/SABG), a lysosomal enzyme detectable at pH = 6.0. However, the proper identification of senescent cells is still insufficient and lacks adequate quantification parameters. In this paper we showed that fractal and gray level co-occurrence matrix (GLCM) texture analysis of cell nuclei are able to discriminate senescent from non-senescent cells. A total of 105 images of DTSCs nuclei from both SA-β-Gal+ and SA-β-Galcells were used in this study. All of the computed parameters (fractal dimension, angular second moment, inverse difference moment, contrast, entropy and correlation) showed significant statistical difference between the nuclei of senescent and non-senescent stem cells (p < 0.001). Our results show the possible practical value of fractal and GLCM texture analysis as the new markers in the determination and quantification of DTSCs senescence. Correspondence to: Nela Puškaš, MD, PhD, Assistant Professor Institute of Histology and Embryology “Aleksandar Đ. Kostić”, School of Medicine, University of Belgrade, Višegradska 26 11000 Belgrade, Serbia, Tel: 00381113607146; Fax: 00381113612567; E-mail: nela@dr.com


Introduction
Dental pulp stem cells (DPSCs) represent a heterogeneous population of multipotent stem cells that have the ability to differentiate into various types of cells [1]. Since they can be easily obtained from teeth during routine dental procedures, especially DPSCs from human deciduous teeth (also known as deciduous teeth stem cells, DTSCs), have gained a very large attention in recent years from a potential therapeutic standpoint [2].
All cells, including stem cells, have a limited lifespan. After a certain number of divisions, stem cells are subjected to aging, which is characterized by different morphofunctional parameters. One of those changes is the accumulation of lysosomal enzyme β-galactosidase at pH = 6.0 (senescence-associated β-galactosidase/SA-β-Gal/SABG), whose activity is associated with cell senescence in vitro [3,4]. The process of cellular senescence refers to irreversible growth arrest, which is followed by changes in nuclear factors [5]. Although SA-β-Gal is widely used as a biomarker for detecting aging and senescence in stem cells in vitro [6], the proper identification of senescent cells is still insufficient and cell-based methods for identification of those cells are not quantitative [7].
Fractal analysis is a method used to measure the complexity of an object. It is widely used in several different areas of biological and medical sciences, such as neuroscience [8][9][10][11], histological diagnosis and quantification of neoplasms [12][13][14][15], blood vessels changes [16,17], assessment of cell aging [18] and many others. New methods and modifications of fractal analysis are also being discovered [19], thus allowing this mathematical method to be applied in different areas of natural sciences. Unlike fractal analysis, gray level co-occurrence matrix (GLCM) texture analysis can be used in biomedical research to analyze texture features of histological images and also to quantify structural changes in cells and tissues [20,21].
Based on this background, our idea was to determine whether fractal and GLCM texture analysis of DTSCs nuclei could prove as a useful tool in the detection and quantification of stem cell senescence in vitro.

Deciduous teeth stem cells isolation
Deciduous teeth stem cells were isolated from a healthy deciduous maxillary central incisor obtained from a 7 year old child. The tooth extraction was done with informed consent of its parents and ethical comity approval. Pulp tissue was minced into 1-2 mm fragments, transferred to 35 mm Petri dishes and cultivated using Dulbecco's modified Eagle's medium (DMEM) / Ham's F12 (1:1, Invitrogen, Carlsbad, California, USA) supplemented with 10% FBS MSC qualified (Invitrogen, Carlsbad, California, USA) and AA solution, and incubated at standard conditions (37°C, 5% CO 2 ). The DTSCs used for this experiment were from passage number four. Additional experimental information about dental pulp and stem cell isolation can be found in Martacić et al. [22].

Senescence-associated β-galactosidase activity
Senescence-associated β-galactosidase activity was determined by using a commercial kit (Senescence Cells Histochemical Staining Kit -Sigma-Aldrich) according to the manufacturer's instructions. Approximately around 2.5 x 10 4 DTSCs were cultured in each well (Multiwell Glass Chamber Slides, BD Biosciences) during 24 hours, after which they were washed in PBS, fixed with a 4% formaldehyde for 5 minutes at room temperature and then incubated at 37°C overnight in a freshly prepared staining solution containing X-Gal (5-Bromo-4chloro-3-indolyl β-D-galactoside) at pH = 6.0. After washing in PBS, the cells were counterstained with Mayer's hematoxylin for better visualization of nuclei. The SA-β-Gal positive (SA-β-Gal + ) and negative (SA-β-Gal -) cells are shown in Figure 1.

Acquisition and graphic processing of images
For fractal and GLCM texture analysis of DTSCs nuclei, digital photos were taken with an Olympus digital camera (Olympus C5060-ADU) and analySIS 5.0 software (Soft Imaging System, Olympus) at x1000 magnification with immersion. The images were saved in TIFF format with a dimension of 2592 x 1944 pixels, resolution of 200 DPI and bit depth of 24. All picture elements, except the nucleus itself, were removed. Also, in order to avoid the influence of rotational variations on the value of fractal dimension (FD) all nuclei were placed centrally and rotated so that their longer axis was parallel to x axis (Figures 2A  and 2E). After this graphic processing, a total of 105 images containing a single nuclei per picture (obtained from 55 SA-β-Gal + and 50 SA-β-Galcells), were used for further fractal and GLCM texture analysis.

Theory/calculation Fractal analysis using the box-counting method
The lower and upper box-dimensions of a subset n F R ⊂ are respectively defined by and if those values are equal then the common value is referred to as the box-counting dimension of F and is denoted by where N δ (F) can be the smallest number of closed balls of radius δ that cover F, the smallest number of cubes of side δ that cover F, the number of δ-mesh cubes that intersect F, the smallest number of sets of diameter at most δ that cover F or the largest number of disjoint balls of radius δ with centers in F [23]. The box-counting method, applied on digitized images, covers the image with a grid of square cells (with cell size r), where for binary images the cell size is expressed as the number of pixels. The number of squares N(r) needed to cover the image is given by a power law where D B is the box dimension, obtained as an absolute value of the slope of the log-log relationship between N(r) and r [24].
Image processing and fractal analysis were carried out in the ImageJ 1.48v software (NIH, Bethesda, USA; free download from http://rsbweb.nih.gov/ij). We calculated two forms of FD: the FD of the "binary-outlined" nucleus (D out ), which measures the shape of an image ( Figures 2C and 2G) and the FD of the corresponding "black-  and-white" image (D bin ), which evaluates the space-filling property of an image ( Figures 2D and 2H).
In order to calculate D out , firstly the images were converted to an 8-bit format (pathway: ImageJ>Image>Type>8-bit), followed by the Find Edges command in order to outline the nucleus (pathway: ImageJ>Process>Find Edges) and then finally transformed to a binary format (pathway: ImageJ>Process>Binary>Make binary) and skeletonized (pathway: ImageJ>Process>Binary>Skeletonize) ( Figure  2). The same procedure was used to calculate D bin , but with an addition of the Fil command after the binarization process, which in this case, represented the last step in the process of image transformation ( Figures 2D and 2H) The fractal analysis was done using the box-counting method (pathway: ImageJ>Analyze>Tools>Fractal Box Count). By default, the box sizes for counting the fractal dimension using the ImageJ software are 2, 3, 4, 6,8,12,16,32,64. In this paper, the box sizes were 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, obtained as an increasing geometric progression 2 n where n = 0, 1, 2…10, as also used in our previous papers [9,19].

Gray level co-occurrence matrix texture analysis
In order to calculate GLCM parameters of DTSCs nuclei, an ImageJ plugin called "Texture Analyzer" version 0.4 was applied (http://rsb. info.nih.gov/ij/plugins/texture.html; developed by Julio E. Cabrera). The analysis was performed on grayscale 8-bit images (Fig. 2B, 2F), with the size of the step in pixels set to 1 and the direction of the step set to 0°. The following five parameters were computed: angular second moment (ASM), inverse difference moment (IDM), contrast (CON), entropy (ENT) and correlation (COR) by following formulas: , The term p(i,j) is the ith and jth entry in a normalized graytone spatial-dependence matrix (i.e. the co-occurrence matrix) and N g is the number of distinct gray levels in the quantized image. The above mentioned formulas, originaly described by Haralick et al. [25] are implemented in the "Texture Analyzer" Image J plugin. The only parameter computed differently is COR, which is calculated as described by Walker et al. [26]. Texture analysis aims to indentify texture features of object in an image by analyzing combinations of intensities at specified positions that are relative to each other in the image [27]. Statistical texture analysis can therefore be classified into first-order, second-order and higher-order statistics according to the number of intensity points (pixels) in each combination. The GLCM texture analysis aims to describe the second order texture features [27]. Originaly Harlick et al., [25] described a total of 14 statistical texture measures that are based on gray tone spatial dependencies. From those 14 parameters, 5 parameters are commonly used in biomedical research for image classification purposes [25]. ASM, also known as, Energy and IDM, also known as Homogeneity, are measures of image homogeneity. High ASM and IDM values are thus present when the image is made of pixels with very similar gray level values, which points to the texture that is uniform and contains repetitive structures [28]. CON and ENT are measures of image heterogeneity, and represent measures of gray level variation and spatial disorder. High CON values can be found in heavy textures images, while high ENT indicates that the image texture is statistically more chaotic [28]. COR is a measure of gray level linear dependence between the pixels at the specified positions relative to each other [27].

Statistical analysis
Statistical analysis was performed using Student's t test and Mann-Whitney's U test, where appropriate. Results are represented as mean ± standard error of mean, and as median with interquartile range. For p values less than 0.05, the differences between the means of populations were considered statistically significant.

Results
The fractal analysis of "binary-outlined" nucleus showed no statistical significance between two groups. On the other hand, the fractal analysis of "black-and-white" nucleus images showed significant statistical differences between two groups of cells (Table 1).
GLCM texture analysis showed high statistical significance in all of the five computed parameters (Table 2). *Results are represented as mean ± standard error of mean. Abbreviations: D out -fractal dimension of the "binary-outlined" nucleus images; D bin -fractal dimension of "black-and-white" nucleus images. Table 1. The average fractal dimension of DTSCs "binary-outlined" and "black-and-white" nuclei obtained from SA-β-Galand SA-β-Gal + groups.

Discussion
Although standard quantitative methods in science are based on classical Euclidean geometry, early research showed that the most common biological patterns were characterized by fractal geometry [29,30]. Up to now the concept of fractal geometry is being used in diverse research areas and is proving to be an increasingly useful tool. In fractal analysis, a type of quantitative analysis derived from fractal geometry [31], there are two basic methods to measure the FD of objects in a plane: the first and most commonly used is a lengthrelated method, while the second method is a mass-related method [29]. Although there are no strict guidelines for the application of fractal analysis, the majority of researches tend to use box-counting method to calculate the FD of digital images, making this method one of the most frequently used [31].
In our study the box-counting method of fractal analysis applied on nuclei of SA-β-Gal + and SA-β-Galcells, i.e. senescent and nonsenescent cells, gave two different types of results. The first result showed no statistical significance between the mean FD values of the "binary-outlined" nucleus, while the second result showed high statistical significance when fractal analysis was applied on "black-andwhite" corresponding images of the nuclei. These contradictory results can be explained by the different parameters which were measured by these two methods. The first method, measured the shape of the nucleus, while the second method evaluated the space-filling property of an image of the nucleus. It has been reported previously [30] that the box-counting method applied on binary images of neurons depends on the way the image is displayed. While the box dimension of "binaryskeletonized" images summarizes the degree of dendrite aberrations from straight lines [32], the same parameter for "binary-outline" images defines the irregularity in the shape of the image. In addition, the box dimension of the "black-and-white" images represents a robust technique for estimating how the extent of the image deviates from the extent of the geometric figure, according to resolution [33]. Thus, according to our results, the fractal analysis of certain geometrical shapes, such as the shape of cell nuclei, should be done on "black-andwhite" images, which in these conditions give more adequate results than the analysis on the "binary-outline" images.
The biological differences obtained in the values of nucleus FD in senescent and non-senescent cells on "black-and-white" images can be explained by changes that occur in the process of stem cell aging [4]. Namely, it was noticed that increase in whole cell size and nuclear size is often associated with senescence [34,35]. Also, abnormal nuclear morphology, such as irregular shape or fragmentation, is often seen in senescent cells [36]. Thus, it is not surprising that "black-and-white" nuclei images of SA-β-Gal + senescent cells have lower values of FD than β-gal-cells, most likely due to the loss of their normal shape and structure. This result is in accordance with other researches that showed that with aging, FD of certain tissue decreases, thus pointing to the loss of structural complexity as a consequence of tissue age-related degradation and deterioration [18].
Besides fractal analysis, GLCM texture analysis has proven to be an even better method in detecting the stem cell senescence in vitro. In our research, all five computed GLCM parameters showed high statistical significance between two groups of cells. As mentioned earlier, ASM and IDM are two parameters that measure tissue homogeneity, whereas second two parameters, CON and ENT measure tissue heterogeneity [25]. Also, IDM relates inversely to CON, while ASM is an opposite of ENT. COR measures the grey-linear dependency of neighboring pixels, where high COR indicates greater pixel similarity and low COR points towards irregularity in image texture [37]. Our results showed that nuclei of SA-β-Gal + cells, have higher values of ENT and CON than SA-β-Galcells, while ASM, IDM and COR of cell nuclei are higher in the SA-β-Galgroup. These results can be explained by the changes in the chromatin structure in senescent cells, whose nuclei show significant structural changes. The most prominent is formation of the senescence-associated heterochromatic foci [38]. Second is distorted organization of the nuclear lamina [5,39], such as downregulation of lamin B1 [40], which leads to the permeabilization of the nuclear envelope. It causes a loss of chromatin from the nucleus to the cytoplasm, together with a progressive decrease in histone content [41]. In this way, the normal nuclear architectonic is lost, leading to an increase in pixel heterogeneity and decrease in pixel homogeneity, which can be detected by GLCM parameters of texture analysis, as it is shown by our results. These results are in accordance with Pantic et al. [42] who have shown an increase in nuclear ENT of hematopoietic erythroid cells with age.
To our knowledge, this is the first study to show the usefulness of fractal and GLCM texture analysis in the detection of stem cell senescence. Our study had certain limitations, due to which it can be considered as a pilot study. The main limitation is the number of biological specimens from which images of DTSCs nuclei were extracted. Although, stem cells were obtained from only one tooth, this is not uncommon in stem cell research, where different experimental procedures are often done on stem cells isolated from one tooth or patient [43][44][45]. Thus, the next research step could be the application of these two methods on a larger number of specimens and on different types of mesenchymal stem cells isolated from various tissues. Besides β-galactosidase staining, additional markers of stem cell senescence should be included in future studies to obtain the most accurate results about the application of these two mathematical methods in senescence quantification.

Conclusion
As shown in our research, fractal and GLCM texture analysis of cell nuclei have proven to be a useful method to quantify stem cell senescence. Besides theoretical, these results also have a practical value, as they may become a new marker for laboratory detection of stem cell senescence, since they are easy to apply and carry no financial cost. The possible role of these methods in future studies can be directed to discriminate stem cells from different passages, to study the effects of different factors on the nuclear complexity of various cell lines or as tool for detecting stem cell differentiation. Although still in the early phase of application, several studies have recently used fractal and GLCM texture analysis for the similar purpose [21,46]. If future studies confirm the validity of these methods, a possible next step in the process of application could be a development of automated software, which would be able to calculate fractal and GLCM texture parameters in real time, without previous image processing.