- Research article
- Open Access
High content image analysis for human H4 neuroglioma cells exposed to CuO nanoparticles
© Li et al; licensee BioMed Central Ltd. 2007
- Received: 05 February 2007
- Accepted: 09 October 2007
- Published: 09 October 2007
High content screening (HCS)-based image analysis is becoming an important and widely used research tool. Capitalizing this technology, ample cellular information can be extracted from the high content cellular images. In this study, an automated, reliable and quantitative cellular image analysis system developed in house has been employed to quantify the toxic responses of human H4 neuroglioma cells exposed to metal oxide nanoparticles. This system has been proved to be an essential tool in our study.
The cellular images of H4 neuroglioma cells exposed to different concentrations of CuO nanoparticles were sampled using IN Cell Analyzer 1000. A fully automated cellular image analysis system has been developed to perform the image analysis for cell viability. A multiple adaptive thresholding method was used to classify the pixels of the nuclei image into three classes: bright nuclei, dark nuclei, and background. During the development of our image analysis methodology, we have achieved the followings: (1) The Gaussian filtering with proper scale has been applied to the cellular images for generation of a local intensity maximum inside each nucleus; (2) a novel local intensity maxima detection method based on the gradient vector field has been established; and (3) a statistical model based splitting method was proposed to overcome the under segmentation problem. Computational results indicate that 95.9% nuclei can be detected and segmented correctly by the proposed image analysis system.
The proposed automated image analysis system can effectively segment the images of human H4 neuroglioma cells exposed to CuO nanoparticles. The computational results confirmed our biological finding that human H4 neuroglioma cells had a dose-dependent toxic response to the insult of CuO nanoparticles.
- Probability Density Function
- Cellular Image
- High Content Screening
- Pixel Classification
- Gradient Vector Field
A precise determination of cell death model is essential for biomedical researches as cell death pathways are intimately associated with normal physiology and disease-related pathogenesis. The widely used colormetric cytotoxicity assays such as lactate dehydrogenase (LDH) release, MTT [3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide]/MTS [3-(4,5-dimethylthiazol-2-yl)-5-(3-carboxymethoxyphenyl)-2-(4-sulfophenyl)-2H-tetrazolium, inner salt] based assays, etc., can only evaluate the viability of cell ensemble. Thus there is a strong demand for sensitive, quantitative, reliable and automated methods for the accurate assessment of cellular proliferation status with high contents of cellular information. As a modern drug discovery tool, high content screening (HCS)  using automated fluorescence microscope is becoming an important and widely used research tool to assist researchers understanding complex cellular processes in disease pathogenesis, drug target validation and drug lead identification [2, 3]. Using the HCS technology, abundant spatial and temporal morphologic information can be extracted from the cellular images, and the information can be used to determine whether a potential drug affects the functions of proteins or genes involved in a disease process. However, it has been a challenge to perform quantitative analysis of the complex cellular images, and this significantly restricts the potential of HCS in drug discovery . Thus, the availability of fully automated cellular image analysis systems is critical to the success of HCS.
In [6, 7] some nuclei segmentation methods were proposed. They combined the intensity gradient information with the shape information to separate the clustered nuclei by using a statistical model to merge the fragments of nuclei. Since the bright nuclei cluster together heavily, and the dark nuclei cannot be accurately separated from the background, these methods tend to fail because the shape information is not accurate. In addition, edge based segmentation methods will fail due to the noisy and discontinuous edges . Thresholding methods cannot separate the clustered nuclei . Moreover, the contours' initialization of the snake and level set methods is much more challenging work [10–12].
Human H4 neuroglioma cells purchased from the ATCC (Manassas, VA) were seeded into 96-well cell culture plates and cultured in Dulbecco's modified Eagle medium (DMEM) supplemented with 10% fetal bovine serum, 1% penicillin-streptomycin solution (Sigma Chemical Co., St. Louis). The cells were incubated for 48 hours under the cell culture conditions (95% O2, 5% CO2, 85% humidity, 37°C), together with CuO nanoparticles at a concentration range of 0.01–100 μM. Then a live/dead assay kit (Molecular probes/Invitrogen) for cell viability was applied to the cells according to the manufacturer's instruction. In brief, the cells were cultured at 37°C for 30 min, with ethidium homodimer-1 (3 μM, for dead cells), and Hoechst dye (16 μM, for nuclear staining) in each well. High-content cellular fluorescence images were acquired using an automatic fluorescence microscope – IN Cell Analyzer 1000 (GE Healthcare). The objective magnification is 10×. Numerical aperture is 0.45, pixel width is 0.645 μm and pixel height is 0.645 μm for all the images taken. The size of each image is 1040 × 1392 pixels.
Choice of parameters
Values and description of the parameters used in the proposed method
Threshold for bright nuclei binarization
Threshold for dark nuclei binarization
Sigma of Gaussian filtering for bright nuclei
Sigma of Gaussian filtering for dark nuclei
Threshold of central point detection of bright nuclei
Threshold of central point detection of dark nuclei
Radius of local suppression of bright nuclei; D b is the diameter of bright nuclei (D b ≈ 9.675 micron).
Radius of local suppression of bright nuclei; D d is the diameter of dark nuclei (D d ≈ 19.35 micron).
Threshold of the PDF score
Robustness test of the parameters: c b , c d , σ b and σ d
# of detected nuclei
# of detected nuclei
# of detected nuclei
# of detected nuclei
Validation and comparison of segmentation
Validation of the proposed method on ten randomly selected nuclei images
# of nuclei (manual counted)
# of nuclei (correctly – segmented)
# of nuclei (over – segmented)
# of nuclei (under – segmented)
# of nuclei (missed)
# of nuclei (noises)
Comparison of segmentation results: Watershed vs. the proposed method
Correctly segmented (%)
Analysis of cell death induced by CuO nanoparticles
P-values of the T tests for H4 cell death rate comparison: the CuO nanoparticles treated vs. untreated
Concentration of CuO nanoparticle (μM)
Herein we present a fully automated cellular image analysis system for quantitative analysis of the viability of human H4 neuroglioma cells exposed to CuO nanoparticles with different concentrations (0.01 – 100 μM). A multiple thresholding method was used to classify nuclei image into three classes: bright nuclei, dark nuclei, and background, based on the background correction algorithm. Following this, a method for fining local image intensity maxima using the Gaussian filtering and gradient vector field was developed to detect the nuclei. A statistical model based splitting method was proposed to reduce the under segmentation problem. The experimental results show that 95.9% nuclei are segmented correctly using the proposed image analysis protocol. Its application on our experimental data sets further indicates that the human H4 neuroglioma cells have a concentration-dependent toxic response to the insult of CuO nanoparticles.
Image pre-processing and pixel classification
No imaging system is perfect, and it is imperative to perform pre-processing to remove the effects of noises, artifacts, uneven illumination, and striped patterns [6, 7, 13, 14] that degrade image quality. To remove the noises and other artifacts without blurring the edges, the median filtering [6, 7] was applied. For uneven illumination and striped patterns, a data driven background correction algorithm [13, 14] was employed to correct the degeneration of the images. The algorithm makes use of the cubic B-splines which have good features, such as continuouity and smoothness, to estimate the background iteratively, and the convergence of this algorithm is fast. Image pre-processing produced images with improved quality.
In this study, the pixel classification means to classify each pixel into the one of three classes: background, dark nuclei and bright nuclei. There are two reasons for doing pixel classification. First, separating the nuclei pixels from the background can reduce the influence of the background in following dark nuclei and bright nuclei detection. Secondly, two kinds of nuclei: bright nuclei and dark nuclei displayed different features in the image, as shown in Figure 1. The bright nuclei, which have high intensity, rice shape, and smaller size, form a tight cluster. The dark nuclei, which have low intensity, round shape and larger size, are scattered. Hence it is reasonable to analyze the bright and dark nuclei separately due to their different attributes.
To achieve pixel classification, we employed the background correction algorithm [13, 14] as a multiple adaptive thresholding method. The basic idea of this method is straightforward. We can visually separate the nuclei from the background into different classes due to the discontinuity of intensity between the background and the two kinds of nuclei. Based on this fact, we can classify one pixel into one of the three classes based on its intensity difference between the real image and the estimated background image obtained by the background correction algorithm. Mathematically the multiple adaptive thresholding method can be written as:
Q(x, y, c) = q(I(x, y) - B(x, y) - c*σ B )
where Cd⊕b(x, y), C d (x, y) and C b (x, y) denote the classes of nucleus (dark and bright), dark nucleus, and bright nucleus, respectively. The noisy fragments were removed based on the size, and the holes on the nuclei objects were considered as the noisy fragments in the negative image. The two thresholds, c d and c b , were obtained experimentally, and we processed all the nuclei images with the same c d and c b . The parameter selection is discussed in more detail in the 'Choice of parameters' Section. In the following sections, we used the dark nuclei image and bright nuclei image to denote the images which only contain the dark nuclei pixels and the bright nuclei pixels, respectively.
Although the nuclei are separated from the background by the multiple adaptive threshold method, many clustered nuclei are under-segmented. To segment the clustered nuclei, the positions of nuclei need to be detected, which serves as the seed points of the seeded watershed segmentation algorithm. In the following, we propose a nuclei detection method using the Gaussian filtering and gradient vector field.
Detection of the central points (nuclei)
where I(x, y) is an image function. It is well known that, in the electric field, the free negative electrons move along the electric field lines and stop at the positive electrodes. In the gradient vector filed, the gradient vector lines point to the local maxima of the filtered images. If we view the local maxima and the detected nuclei pixels as the positive electrodes and the free negative electrons respectively, by the same analogy, the nuclei pixels of a nucleus will move along the gradient vector lines in the gradient vector field and at last stop at the central point inside the nucleus. Therefore, these central points will be covered by a number of nuclei pixels whereas the non central points have no one pixel stops at them. Based on this fact, we let the detected nuclei pixels move along the gradient vector lines first, and then the central points can be detected by finding the points which are covered by a significant number of pixels. The motion of pixels along with the gradient vector lines can be achieved as follows: given a pixel (x0, y0), let it move along the direction of the gradient vector in point (x0, y0) to its nearest neighbour (x1, y1), and then pixel (x0, y0) moves again along the direction of the gradient vector in point (x1, y1) to the next nearest point (x2, y2). Repeating this process, pixel (x0, y0) at last will stop at a local maximum. In these detected central points, some noises and redundant (more than one central points appearing inside a single nucleus) central points exist. To suppress the noises, we removed the central points with convergent pixels less than a certain number, T b , for the bright nuclei central points, and T d for the dark nuclei central points. We applied the following criterion to reduce the redundant central points: if the distance between two central points is less than a threshold, r b , for the bright nuclei central points, and r d for the dark nuclei central points, the one with fewer convergent pixels is removed. Finally the detection results of bright and dark nuclei were pooled together. Figure 3-(a) and 3-(b) show the detection results of Figure 9-(a) and 9-(d). Figure 3-(c) shows the detection result of the Figure 1.
Statistical model based splitting method for refining the nuclei detection
Snake model, level-set and seeded watershed methods are a few popular segmentation techniques. However, the snake models need the initial contours near to the true boundaries; the level set method has high computational expenses. Here we employed the seeded watershed based region growing algorithm to segment the nuclei. Figure 10 shows the initial segmented result of Figure 1.
Gaussian Probability density function (PDF) model
Features used in the PDF model
Major Axis Length
Minor Axis Length
Standard deviation of intensity
Splitting under-segmented nuclei
After obtaining two new nuclei via the splitting step, it is assumed that the PDF scores of the two new generated nuclei should be greater than the original one. Thus the following criterion is established for splitting under-segmented nuclei: if P(x c ) ⟨ P(xc2) and P(x c ) ⟨ P(xc2), we accept the splitting result; otherwise, we reject the splitting result. The new nuclei obtained from the splitting step are measured by the PDF model again, and the nuclei whose PDF values are less than the given threshold, T pdf , are sent to the splitting step again. This process is repeated until no new nucleus is generated. Figure 4 presents the final segmentation result. The software of the proposed system is available, see Additional file 1.
This work was supported by the Center for Bioinformatics Program grant (to STC Wong) of Harvard Center of Neurodegeneration & Repair, Harvard Medical School, Boston, MA, USA. X. Huang is supported by NIH Career Development grant (5K01MH002001) and funds from Radiology Department of Brigham and Women's Hospital. J. Zhu is supported by NIH Neuroimaging Neuroinformatics Training Program (5K12MH069281-04 to DN Kennedy).
- Perlman Z, Slack M, Feng Y, Mitchison T, Wu L, Altschuler S: Multidimensional drug profiling by automated microscopy. Science. 2004, 306: 1194-1198. 10.1126/science.1100709.View ArticleGoogle Scholar
- Zhou X, Wong S: High content cellular imaging for drug development. IEEE Signal Processing Magazine. 2006, 23: 170-174. 10.1109/MSP.2006.1598095.View ArticleGoogle Scholar
- Zhou X, Wong S: Informatics challenges of high-throughput microscopy. IEEE Signal Processing Magazine. 2006, 23: 63 -672. 10.1109/MSP.2006.1628879.View ArticleGoogle Scholar
- Adiga U, Malladi R, Fernandez-Gonzalez R, de Solorzano CO: High-throughput analysis of multispectral images of breast cancer tissue. IEEE Transactions on Image Processing. 2006, 15: 2259-2268. 10.1109/TIP.2006.875205.View ArticleGoogle Scholar
- Nattkemper T: Automatic segmentation of digital micrographs: a survey. Medical Informatics. 2004, 11: 847-851.Google Scholar
- Lin G, Adiga U, Olson K, Guzowski J, Barnes C, Roysam B: A hybrid 3-D watershed algorithm incorporating gradient cues and object models for automatic segmentation of nuclei in confocal image stacks. Cytometry A. 2003, 56: 23-36. 10.1002/cyto.a.10079.View ArticleGoogle Scholar
- Lin G, Chawla M, Olson K, Guzowski J, Barnes C, Roysam B: Hierarchical, model-based merging of multiple fragments for improved three-dimensional segmentation of nuclei. Cytometry A. 2005, 63: 20-33.View ArticleGoogle Scholar
- Garbay C: Image structure representation and processing discussion of some segmentation methods in cytology. IEEE Transactions on Pattern Analysis and Machine Intelligence. 1986, 8 (2): 140-146.View ArticleGoogle Scholar
- Kohler R: A Segmentation System Based on Thresholding. Compter Graphics and Image Processing. 1981, 15: 319-337.View ArticleGoogle Scholar
- Osher S, Sethian J: Fronts propagating with curvature dependent speed: Algorithms based on the hamilton-jacobi formalism. Journal of Computational Physics. 1988, 79: 12-49. 10.1016/0021-9991(88)90002-2.View ArticleGoogle Scholar
- Sethian J: Level Set Methods: Evolving interfaces in geometry, fluid mechanics, computer vision, and material science. 1996, Cambridge University PressGoogle Scholar
- Chen X, Zhou X, Wong S: Automated segmentation, classification, and tracking of cancer cell nuclei in time-lapse microscopy. IEEE Transactions on Biomedical Enineering. 2006, 53 (4): 762-766. 10.1109/TBME.2006.870201.View ArticleGoogle Scholar
- Lindblad J, Wahlby C, Bengtsson E, Zaltsman A: Image analysis for automatic segmentation of cytoplasms and classification of Rac1 activation. Cytometry A. 2004, 57 (1): 22-33. 10.1002/cyto.a.10107.View ArticleGoogle Scholar
- Wahlby C, Lindblad J, Vondrus M, Bengtsson E, Bjorkesten L: Algorithms for cytoplasm segmentation of fluorescence labelled cells. Anal Cell Pathol. 2002, 24 (2-3): 101-111.View ArticleGoogle Scholar
- Beucher S: The watershed transformation applied to image segmentation. Scanning Microscopy International. 1992, 6: 299-314.Google Scholar
- Vincent L, Soille P: Watersheds in digital spaces: an efficient algorithm based on immersion simulations. IEEE Transactions on Pattern Analysis and Machine Intelligence. 1991, 13: 583-598. 10.1109/34.87344.View ArticleGoogle Scholar
- CellProfiler. [http://www.cellprofiler.org/]
- Benson JM, Henderson RF, Pickrell JA: Comparative in vitro cytotoxicity of nickel oxides and nickel-copper oxides to rat, mouse, dog pulmonary alveolar macrophages. Journal of toxicology and environmental health. 1988, 24 (3): 373-383.View ArticleGoogle Scholar
- Wang J, Zhou X, Li F, Wong S: Classify Cellular Phenotype in High-Throughput Fluorescence Microcopy Images for RNAi Genome-Wide Screening. IEEE/NLM Life Science Systems & Applications Workshop. 2006, Bethesda, MDGoogle Scholar
- Pudil P, Novovičová J, Kittler J: Floating search methods in feature selection. Pattern Recognition Letters. 1994, 15: 1119-1125. 10.1016/0167-8655(94)90127-9.View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.