Automatic quantitative analysis of morphology of apoptotic HL-60 cells

Abstract

Morphological identification is a widespread procedure to assess the presence of apoptosis by visual inspection of the morphological characteristics or the fluorescence images. The procedure is lengthy and results are observer dependent. A quantitative automatic analysis is objective and would greatly help the routine work. We developed an image processing and segmentation method which combined the Otsu thresholding and morphological operators for apoptosis study. An automatic determination method of apoptotic stages of HL-60 cells with fluorescence images was developed. Comparison was made between normal cells, early apoptotic cells and late apoptotic cells about their geometric parameters which were defined to describe the features of cell morphology. The results demonstrated that the parameters we chose are very representative of the morphological characteristics of apoptotic cells. Significant differences exist between the cells in different stages, and automatic quantification of the differences can be achieved.

Introduction

Apoptosis occurs in an orderly fashion with morphological events including cell shrinkage, the degradation of DNA, orderly condensation of the nucleus, and the production of membrane “blebbs” which are phagocytosed by surrounding cells. This sequence of events was termed “apoptosis” by Kerr et al.(1972[10]). Study of apoptosis has attracted significant attention in cancer research and anticancer technology development in recent years. Much research has been conducted with visual inspection of fluorescent images of stained cells under microscopes (Van Cruchten and van den Broeck, 2002[16]; Kalinichenko and Matveeva, 2008[9]). The morphological identification of apoptotic cells in fluorescent microscopic images is currently the most widely used method, but it is ambiguous because of the limited qualitative description. The technique is also labor intensive, time-consuming, and the accuracy of the final results ultimately depends on the skill and experience of the observer. It has been challenged by practices in analyzing large amount of cells to obtain statistically meaningful results (Hoyt et al., 1997[7]). An automatic approach is therefore needed in the study of morphological changes associated with cell apoptosis.

Most of the published studies focused on segmentation approaches (Gurcan et al. 2006[6]; Dorini et al., 2007[4]; Cseke, 1992[2]) and did not include the extraction or analysis of geometric parameters and quantification of apoptotic cell morphology. Our research aims at quantitatively analyzing the morphological changes in different apoptotic stages of HL-60 cells. For this purpose, we have implemented an image processing and segmentation method, defined adequate parameters and an automatic method for morphological measurement which are presented in the next section followed with experimental results and discussion.

Materials and Methods

Image acquisition

Brief exposure of cultured cells to hydrogen peroxide (H₂O₂) can induce cellular death that proceeds via an apoptotic pathway (Kovar et al., 1997[11]). In this study, human leukemia cells (HL-60) were cultured at 37 °C in a humidified 5 % CO₂ atmosphere in the media of RPMI1640 (Life Technologies, New York) with 10 % heat-inactivated fetal bovine serum (FBS, Life Technologies) and treated with 1.5 mM H₂O₂ for 24 h. They were stained with Hoechst33342 and AnnexinV-FITC/PI (Life Technologies) to determine the apoptosis stage. Cell nuclei marked by PI are identified as late apoptosis in this study (including nerocrosis) which are shown in red color in the fluorescent images. AnnexinV can be conjugated to a variety of fluorochromes, and binds to “flipped” Phosphatidylserine (PS) residues on extracellular membrane leaflet of apoptotic cells. Hoechst33342 is a highly permeable DNA binding dye that enters cells and labels their chromatin regardless of their viability state (Telford et al., 2011[15]). In other words, normal cells are stained by Hoechst33342 (Hoechst33342+, AnnexinV-, PI-), early apoptotic cells are stained by AnnexinV but not PI (AnnexinV+, PI-), and only late apoptotic cells are stained by PI (PI+).

The stained cells in the experiments were photographed in a light channel (Figure 1a(Fig. 1)) and three fluorescent channels (Figure 2a(Fig. 2)) with an inverted fluorescence microscope (IVFM, IX71, Olympus). The experiments were repeated three times independently. The image resolution is 1367×1038 and the calibration setting for the cell images is 32×magnification, which gives 0.3225 µm/pixel.

Figure 1. Light channel images and the segmentation. (a) original image; (b) after pre-processing; (c) segmentation with the Otsu method; (d) after processing with dilation and erosion operators; (e) outlined cells on the original image.

Figure 2. Fluorescent image segmentation. (a) original image; (b) B channel image; (c) B channel image after segmentation; (d) G channel image; (e) G channel image after segmentation; (f) R channel image; (g) R channel image after segmentation.

Segmentation

Automatic and quantitative analysis of cellular morphological changes in fluorescent microscopic images requires the development of algorithms for segmenting the cells individually and for quantifying their shape parameters accurately without manual interaction (Angulo and Schaack, 2008[1]). Thresholding becomes an effective tool of separating objects from backgrounds when their gray levels are substantially different. The morphological method is also widely used in both subsequent processing steps and initial segmentation (Meijering, 2012[12]). The algorithm used in our method combines these methods to achieve faster segmentation for images with different features.

Ideal images for the automatic segmentation should contain compact cells in a uniform background. Raw images obtained in experiments contain noise from samples, the microscope light channel and three fluorescent channels, and need to be pre-processed. The image enhancement algorithm we used to facilitate segmentation is described as follows (Nasonov and Krylov, 2012[13]). For light microscope images, the morphological opening method was first applied to correct non-uniform illumination. Then we used a 5×5 window median filtering combined with grayscale stretch to improve the image quality of light channel and each fluorescent channel of microscopic images. Median filtering is useful for reducing random noise and periodic patterns (Weber and Albrecht, 1997[17]; Huang et al., 1979[8]), and grayscale stretch can expand the dynamic range of a region of interest. The background noise and the interference of other channels were excluded after the pre-processing (Figure 1b(Fig. 1)).

The classical clustering-based Otsu threshold segmentation algorithm is an automatic method to find the threshold which can divide the given image into foreground and background through maximizing the between-class variance (Poletti et al., 2012[14]; Xu et al., 2011[18]). Suppose that pixels in a given image are represented in m gray levels, the ith level has n_i pixels, and the total number of pixels is. The probability of occurrence of each level in the given image isp_i=n_i/N, and the mean value is . Pixels are now divided into two classes C0={1~k} and C1={k+1~m} with probability of and . The mean level of each class is

.

To reach the largest difference as expressed in Eq. 2, the variance between the two classes should reach the maximum. The threshold is given in Eq. 3.

The Otsu method is the least time-consuming segmentation method suitable for the process of a large amount of images, yet very sensitive to noise and the target size. To alleviate the noise impact to the segmented results, we used robust pre-processing methods as described above to enhance the image.

The dilation and erosion operators were repeated in turn as post-seg processing to deal with some small regions and holes in the binarized image (Yua et al. 2009[19]; Di Ruberto et al., 2002[3]). The small objects whose areas were smaller than the averaged size of the objects were removed. The objects on the board were also deleted in case they were not complete (Gonzalez et al., 2009[5]). Figure1d(Fig. 1) and 1e(Fig. 1) show the final segmentation result of a light channel image, and Figure 2b-g(Fig. 2) show images of each channel, as well as their segmentation results.

Parameter extraction and statistical analysis

The cells in light channel images were matched to corresponding cell nuclei based on their location according to a pre-calibrated distance from a cell center to its nucleus center (less than 20 pixels in our study). For each cell, we programmed to search for a nucleus satisfying this condition and identified it as the corresponding nucleus. After this step, the algorithm was designed to determine apoptotic stage of cells, i. e., normal cells (Hoechst33342+, AnnexinV-, PI-), early apoptotic cells (AnnexinV+, PI-) and late apoptotic cells (PI+) according to the staining results of corresponding cell nuclei. This procedure aims at counting cells in each stage. As shown in Figure 3e-g(Fig. 3), cell nuclei of each stage were divided and shown in separated images. Figure 3h(Fig. 3) shows the cell stage determination results (3 normal cell nuclei, 1 late apoptotic cell nucleus and 1 early apoptotic nucleus), which is consistent with that of visual examination. More figures of light channel images and the corresponding fluorescent images with segmented contours are shown in the supplementary material.

Figure 3. Cell stage determination. (a) original light channel image; (b) segmented light channel image; (c) outlined cells on the original light channel image; (d) original fluorescent image; (e) normal.

Several parameters were chosen and defined for the description of HL-60 cells (Table 1(Tab. 1)), including basic and widely used geometric parameters such as cell Area and Perimeter. Shape Factor is also known as an important indicator of cell morphology. The closer to a circular shape the object is, the nearer to 1 the Shape Factor will be. It is related to “cell shrink” or the smoothness of the cell edge. Center Distance between nucleus and cytoplasm was used as a reflection of chromatin margination. We also proposed Smoothness Index (defined as the perimeter of the object divided by the perimeter of an equivalent circle) and Number of Pit Points of the cell membrane as characteristic parameters to represent the extent of “cell blebbing”. The first 5 parameters listed in Table 1(Tab. 1) were also used to describe the morphology of the cell nuclei.

Table 1. Parameters for description of apoptotic cell morphology¹.

For the calculation of Number of Pit Points of cell membrane, the convex hull of each cell membrane was generated. Then the concave area of this cell membrane was found out by removing the cell area from the whole area of the convex hull. Points on the boundary of the concave area reaching the local minimum distances to the centroid of the cell were marked as pit points. Other parameters were calculated according to their definition. All of the algorithms are programmed with MATLAB R2012.

A one-way analysis of variance (one-way ANOVA) test was used in the statistical analysis of the parameters characterizing cell morphology change between the three cell stages (SigmaPlot12.0, San Jose, California).

Results and Discussion

Figure 4(Fig. 4) shows the representative cells for each stage (normal, early apoptosis and late apoptosis) segmented with the algorithm. The morphological parameters (listed in Table 1(Tab. 1)) of (a), (b), (c) and (d) in Figure 4 were calculated for pre-analysis.

Figure 4. Boundary extraction of cell membrane and nucleus. (a) Normal cell; (b) Early apoptotic cell; (c) Early apoptotic cell with blebbing; (d) Late apoptotic cell.

The ones with large variance that can articulate the attribute clearly were chosen for the quantitative analysis (Table 2(Tab. 2)). The parameters of “Number of Pit Points” of the blebbing apoptotic cells and the late apoptotic cells are larger than those of the normal cells and the apoptotic cells of non-blebbing. The Smoothness Index of the blebbing apoptotic cells is the largest, revealing the situation of “cell blebbing”.

Table 2. Parameters of the cells in Figure 4.

Results of the statistical analysis for the three groups of the parameters obtained from the three independent experiments and characterizing the morphology of the HL-60 cells in three stages are summarized in Table 3(Tab. 3). After calculating the mean value and standard deviation (SD) of each parameter, one-way ANOVA tests for the parameters between the three cell stages were conducted. Parameters such as Shape Factor and Roundness of membrane as well as those of nucleus share the same phenomenen as shown in Figure 5a(Fig. 5) for Shape Factor of membrane, with mean values of normal cells significantly larger than those of apoptotic cells. This reveals that normal cells are closer to circles, while early apoptotic cell membranes have more irregular shapes because of “blebbing” and “shrinking”. On the contrary, Perimeter, Ovality, Smoothness Index, Number of Pit Points of membranes, Area, Perimeter, Ovality, Smoothness Index of nucleus and Center Distance between nucleus and cytoplasm show the differences between the three stages as illustrated in Figure 5b(Fig. 5) for Ovality of nucleus. Ovality of nucleus varies greatly between cells in different stages, from which we can see that nucleus shapes of early apoptotic cells are closer to oval than those of cells in the other two stages. One-way ANOVA test results for the Center Distance between nucleus and cytoplasm show that there is a statistically significant difference (P=0.023), which reveals that the margination of cell nucleus can be reflected by the change of this parameter. Its mean value for early apoptotic cells is larger, because chromatin margination usually happens in this stage. In each group, SDs of the parameters of normal cells are the smallest, and those of early apoptotic cells are the largest, revealing that normal cells have relatively homogeneous morphology while apoptotic ones vary greatly.

Table 3. Table 3: Classification results of the three groups.

Figure 5. Statistical results of the parameters in different cell stage. Each column represents a parameter’s mean and standard deviation (SD) for the cells in each stage from the three independent experimental studies. (a) Shape Factor of membrane of normal cells is the largest among the 3 stages (p<0.05 in one-way ANOVA test). (b) Ovality of nucleus of early apoptotic cells is significantly larger than that of normal cells and late apoptotic cells (p<0.05 in one-way ANOVA test).

Conclusion

The experimental results demonstrate that the image processing algorithm presented in this manuscript can effectively extract morphological features of cells from fluorescent microscopic images and achieve accurate staging results. It is less time-consuming as compared to manual contour tracing and parameter extraction. The chosen parameters are representative and can describe the characteristics of cell images precisely. Most of the parameters show great differences between the three stages, which indicates that the morphological differences between normal cells and apoptotic cells can be detected quantitatively with the automatic method.

There are still challenges in developing methods that are applicable and effective for a wide range of cell images with various morphological cellular features and different image characteristics. In addition, further study is required to find out if the parameters chosen in this study can also be applied to other cell types in quantitatively describing apoptosis morphology.

Acknowledgements

The authors acknowledge the support of NSFC (grants #81171342, #81041107, #31000784 and #81201754).

Conflict of interest

The authors declare that there is no conflict of interests regarding the publication of this article.