The segmentation and quantification of cell nuclei are two very significant tasks in the analysis of histological images. Accurate results of cell nuclei segmentation are often adapted to a variety of applications such as the detection of cancerous cell nuclei and the observation of overlapping cellular events occurring during wound healing process in the human body. In this paper, an automated entropy-based thresholding system for segmentation and quantification of cell nuclei from histologically stained images has been presented. The proposed translational computation system aims to integrate clinical insight and computational analysis by identifying and segmenting objects of interest within histological images. Objects of interest and background regions are automatically distinguished by dynamically determining 3 optimal threshold values for the 3 color components of an input image. The threshold values are determined by means of entropy computations that are based on probability distributions of the color intensities of pixels and the spatial similarity of pixel intensities within neighborhoods. The effectiveness of the proposed system was tested over 21 histologically stained images containing approximately 1800 cell nuclei, and the overall performance of the algorithm was found to be promising, with high accuracy and precision values.
Analysis of microscopy images is one of the most fundamental goals in the realm of immunohistochemistry. The primary tasks involved in the analysis of histologically stained tissue sections are cell nuclei counting, detecting abnormal cell nuclei, and the presence of antigens within the target cells. Results derived from these analyses are most frequently used in the clinical setting to help diagnose a wide spectrum of pathologies. In the past, pathologists accomplished most of these tasks by the means of manual measurements; for example, the quantification of total cells and abnormal cells was performed through manual hand counting. These manual methods are not only time consuming, but the results they yield are often susceptible to inconsistency due to human error. However, as the result of recent advancements in microscopic imaging technology and computational image processing techniques [
In recent years, numerous image processing techniques have been proposed for cell nuclei segmentation [
A popular technique in the realm of image processing known as region growing is combined with a graph-cuts-based algorithm that incorporates Laplacian of Gaussian (LoG) filtering to detect cell nuclei [
The selection of the above publications exemplifies the wide range of image processing techniques and practical diagnostics applications that encompass the realm of cell nuclei segmentation in immunohistochemistry. The well-established cell nuclei segmentation and differential immunostaining techniques that have proven to be so valuable in the cancer field are now being applied to the field of wound healing research [
The vast amount of research in this realm also emphasizes on the need for automated computational systems for cell segmentation techniques that produce accurate and reproducible results. However, the task of cell segmentation is still one of the most challenging tasks in biomedical image processing mainly because the histological specimens that are used for the image acquisition process are 2-dimensional sections of 3-dimensional tissue samples [
Thresholding techniques are fairly simple but still effective; they are widely used for the segmentation of histological images because the regions of interest within these images are distinguishable from the other components by visual features such as color and texture [
The following section explains the image acquisition and histological procedures involved in the preparation of the testing dataset. Section
The dataset used in the testing of the proposed algorithm consists of 21 immunohistochemically stained images. The images in the dataset were acquired from human tissue sections derived from PTFE (expanded polytetrafluoroethylene) tubes that were removed at 5, 7, and 14 days after implantation [
The tissue collection and histological staining procedures involved in the preparation of the data set are as follows. Using alcohol and povidone-iodine topical antiseptic, the site of implantation was sterilized and anesthetized using 3 cc lidocaine (1%) without epinephrine. Five, 6.0 cm, of high-porosity PTFE (polytetrafluoroethylene Custom Profile Extrusions, Tempe, AZ) tubes were implanted subcutaneously into the inner aspect of the upper arms of a healthy volunteer subject. Standardized placement was made by a 5.5 cm cannulation of the subcutaneous tissue in a proximal direction. Using a sterile 14-gauge trochar containing PTFE tubing, the skin was punctured, and the trochar was inserted subcutaneously arising through the skin 5.5–6.0 cm away. The trochar was then removed, and the proximal and distal ends of the PTFE tubing were sutured to the skin using a single 5.0 nylon suture. The implantation site was covered with antibiotic ointment and a transparent surgical dressing. On day 14, the PTFE tube was removed and stored in 10% formalin. The wound tissue contained within the fixed PTFE tube was then processed and embedded in paraffin, and 5 micron sections were prepared using standardized histologic techniques. Positive and negative control sections were included to ensure reproducible staining. Hematoxylin & Eosin (H&E) stain was used to highlight the cellular components, and standard immunostaining techniques were used to identify endothelial cells (CD-31), macrophages (CD-68), and contractile cells (
The derived tissue sections were examined using a Zeiss LSM 510 NLO Meta confocal/multiphoton laser scanning microscope. For confocal imaging, the 488, 561, and 633 nm laser lines were used for sample imaging. Images were collected using sequential illumination (i.e., one laser per channel) to avoid signal cross-talk amongst channels. The images were collected using a 63x/1.4 n.a. oil immersion lens (for single photon confocal imaging) or a 63x/1.2 n.a. IR water immersion objective (for multiphoton imaging). The human study was carried out under the approval of the Institutional Review Board of Virginia Commonwealth University, School of Medicine (IRB number 11087).
This section provides an in-depth explanation of the proposed entropy-based image segmentation technique. The flowchart in Figure
Overview of the proposed algorithm.
The preprocessing of an immunohistochemically stained input image
Although there are several options with respect to color spaces wherein the processing of the image can be performed, for this project the cell extraction is primarily performed in the RGB color space. Other color spaces such as YCbCr, LAB, and HSV were tested, and in comparison the RGB color space consistently provided the best results. This is because the objective is to extract cell structures based on the color information present within the image, hence distinguishing the different biological objects within the image. The background removal process starts by separating the RGB color image
Results from background removal process. (a), (c) Input images. (b), (d) Resulting images composed of
A popular histogram equalization technique called Contrast Limited Adaptive Histogram Equalization (CLAHE) is then used in its orignal form to enhance the local contrast in the color component images
In CLAHE, contrast enhancement is performed locally in small regions called “tiles”, each tile’s histogram is equalized to provide a better overall visual distinction between target objects (cell nuclei) and background (intercellular matter). Additionally, the use of CLAHE ensures that the stained cell nuclei in the tissue section are enhanced uniformly, thereby providing accurate recognition of cell nuclei irrespective of the influences of different staining procedures. The histograms derived from the operation of CLAHE are chosen to maintain a uniform shape. The computation of entropy in the proposed algorithm is based on Shannon’s entropy, for which the net information values are calculated within 5-by-5 neighbourhoods throughout the image; the number of tiles for CLAHE’s operation is chosen to be close to the total number of 5-by-5 neighbourhoods present in the input image. Once the contrast enhancement is performed on
After CLAHE is performed on the 3 color component images, the thresholding technique described below is applied to each component image. The proposed entropy-based thresholding method has 3 steps: computation of the color component level spatial correlation (CCLSC) histogram, computation of object and background probabilities, and the computation of object and background entropies. Threshold values for each color component image are obtained once all the calculations are performed. The mathematical computations involved in each step are described in the following sections.
The entropy-based thresholding technique relies on the CCLSC histogram which is a modified version of the Grey-level spatial correlation histogram presented in [
Let
The similarity index is then computed within every pixel neighbourhood, and the degree of similarity is based on the difference of intensity values between the pixel located at the center of the
The similarity index
Surface plot of the CCLSC histogram.
The calculation of object and background entropies require probability distributions associated with an image’s object and background regions that are derived in the following way. A threshold value
The probability distributions described in the previous section are used to compute the object and background entropies. According to the principle of Shannon’s entropy, the measure of uncertainty from a source equals the net value of information obtained from the source. Features such as noise and edges are associated with higher entropy values because they produce discontinuities between the object and the background which produce more uncertainity in images, that is, net information. As noted in Section
Graph of weight function
The object entropy
After the computation of entropies the function
Performing the procedures described in previous section on the preprocessed images
It was experimentally observed that the pixels representing cell nuclei in histologically stained images are composed of lower intensity values in the red and green color component images and higher intensity values in the blue component color image. Therefore, all pixels in
(a) Input image
The image
The second step in the postprocessing of image
(a) Input image
In order to quantify the cell nuclei, the postprocessed image
(a) Input image
The results obtained by testing the proposed automated segmentation technique on a dataset of 21 immunohistochemically stained images are presented in Table
Results of automated segmentation performance on 21 test images.
Stain | ID/magnification | Manual quantification (cells) | Automated quantification (cells) | Accuracy (%) | Precision (%) | Sensitivity (%) | Specificity (%) |
---|---|---|---|---|---|---|---|
SMActin | Day 5–40x | 101 | 100 | 92.96 | 91 | 90.01 | 94.67 |
Day 5–60x | 59 | 60 | 91.82 | 88.33 | 89.83 | 93 | |
Day 7–40x | 166 | 187 | 93.55 | 85.56 | 96.38 | 92.19 | |
Day 7–60x | 87 | 94 | 93.96 | 89.24 | 95.40 | 93.10 | |
Day 14–40x | 85 | 88 | 92.75 | 89.77 | 92.94 | 92.62 | |
| |||||||
H&E | Day 5–40x | 86 | 86 | 100 | 100 | 100 | 100 |
Day 7–40x | 141 | 138 | 99.05 | 100 | 97.87 | 100 | |
Day 7–60x | 66 | 71 | 95.10 | 90.14 | 96.96 | 94.06 | |
Day 14–40x | 105 | 108 | 98.69 | 97.22 | 100 | 97.6 | |
Day 14–60x | 42 | 38 | 95.28 | 100 | 88.095 | 100 | |
| |||||||
CD-31 | Day 5–40x | 76 | 77 | 99.38 | 98.70 | 100 | 98.85 |
Day 5–60x | 52 | 49 | 97.43 | 100 | 94.23 | 100 | |
Day 7–40x | 140 | 188 | 94.32 | 77.77 | 97.22 | 93.66 | |
Day 7–60x | 61 | 59 | 97.61 | 100 | 95.08 | 100 | |
Day 14–40x | 98 | 104 | 97.97 | 94.23 | 100 | 97.97 | |
Day 14–60x | 67 | 67 | 100 | 100 | 100 | 100 | |
| |||||||
CD-68 | Day 5–40x | 99 | 96 | 91.36 | 76.19 | 96.96 | 89.39 |
Day 5–60x | 68 | 66 | 95.13 | 84.61 | 97.05 | 94.54 | |
Day 14–40x | 143 | 138 | 98.55 | 100 | 96.50 | 100 | |
Day 14–60x | 69 | 70 | 99.49 | 98.57 | 100 | 99.23 | |
| |||||||
Total | 1811 | 1884 | |||||
Average | 95.55 | 91.27 | 96.45 | 95.07 |
The testing dataset consisted of 21 images belonging to a single patient that were stained using either Hematoxylin & Eosin (H&E) stain, cluster of differentiation 31 (CD-31), cluster of differentiation 68 (CD-68), or alpha-smooth muscle actin (
In comparison to a related method for cell segmentation based on shape stability [
A novel translational computation system for automated cell nuclei segmentation and quantification has been proposed in this paper. Cell nuclei segmentation is a task that has several medical motivations ranging from the detection of malignant cell nuclei (tumor) in cancerous tissue images to the observation of cell nuclei for the characterization of the wound healing process within the human body. The proposed system uses an entropy-based thresholding technique to yield 3 optimal threshold values that are used to segment cell nuclei from images in the RGB color space. The entropy-based computations were based of the concepts introduced in [
Virginia Commonwealth University Reanimation Engineering Science Center (VCURES) and the Biomedical Signal and Image Processing Lab of the Computer Science Department at Virginia Commonwealth University have supported the work presented in this paper.