1. Introduction

Microarray technology came on time to cover the need to monitor in parallel all the DNA sequences and to have the adequate sensibility to detect the variation of gene expression. There may be tens of thousands of spots on an array. Each spot contains tens of millions of identical DNA molecules with lengths from tens to hundreds of nucleotides. Afterwards, the microarray slide is exposed to a set of labeled cDNA samples, which are derived from tissue of interest. With the completion of hybridization reaction, the amount of the target that bounds to each sample is measured with the aid of image capturing devices and computer technology. The measurement is based on the intensity of the spot. There are three basic steps in the processing of microarray image [1]; gridding, segmentation, and quantification. Image segmentation is the process of distinguishing objects from their background [2]. It is usually the first step in vision systems, and is the basis for further processing such as description or recognition. The goal of segmentation is to extract important features from images. Segmentation of an image can also be seen, in practice, as the classification of each image pixel to be assigned to one of the image compositions.

Different segmentation methods have been presented include the dynamic system modeling based approach [3] performs pixel clustering operations in a parallel manner to speed-up the segmentation process. The cellular neural network scheme [4, 5] segments the spots by performing a number of operations such as background clean-up, grid analysis, irregular spot elimination, and intensity analysis. The morphology based approach [6] uses a series of optional steps to segment the microarray image. The combination of Markov random field based grid segmentation and active contour modeling constitutes an approach suitable for spot detection and segmentation [7]. The two-stage clustering based approach [8] is comprised of spots boundaries adjusting and intensity-based partitioning operations. The use of adaptive thresholding and statistical intensity modeling is the base for some segmentation schemes [9], whereas another approach [10] uses a seeded region growing algorithm to identify spots of different shapes and sizes. Histogram and thresholding operations were used to classify microarray image samples into either foreground (spots) or background pixels [11].

One of the key steps in extracting information from a microarray image is the segmentation whose aim is to identify which pixels within an image represent which gene. This task is greatly complicated by noise within the image and a wide degree of variation in the values of the pixels belonging to a typical spot. A proposed segmentation method is presented on different noisy microarray images, which gives high accuracy. The paper is organized as follows: a brief introduction is presented in this section, section 2 presents the used materials, section 3 summarizes the proposed segmentation method applied on different cDNA microarray noisy image and section 4 discusses the results of the applied algorithm on microarray data set images. Conclusions are presented in section 5.

2. Materials

Twenty different images have been selected from two different data sets, to test the performance of the proposed methods. They have different scanning resolutions, and different noise types, in order to study the flexibility of the proposed methods to detect spots with different sizes and features. The images are stored in TIFF files with 16-bit gray level depth. A chosen microarray image with various kinds of artifacts or noises is drawn from Princeton University Microarray (PUMA) database [12].

The data of the used images are described as follows:

The first image includes thirty-two sub-grids representing acute lymphoblastic leukemia tissues (PUMA Experiment ID: 10223). The slide name is (shae082), and it is a cDNA microarray spotted by a total of 24192 genes.

The second image includes thirty-two sub-grids representing acute lymphoblastic leukemia tissues (PUMA Experiment ID: 10224). The slide name is (shae083), and it is a cDNA microarray spotted by a total of 24192 genes.

The third image includes thirty-two sub-grids representing acute lymphoblastic leukemia tissues (PUMA Experiment ID: 10966). The slide name is (shae086), and it is a cDNA microarray spotted by a total of 24192 genes.

MATLAB [13] is used for data analysis and technical computing, as it is a high performance and powerful tool. MATLAB version is 6.5 on a windows XP platform. The P.C used has a processor: Intel Core r2 4300 CPU with 2GB of memory.

3. Methods

Once grids have been placed [14], discrimination between areas that are considered the spot signal and areas that are considered the background signal must be carried out. The process, by which each individual cell in the grid must be selected to determine the spot signal and to estimate the background hybridization, is called segmentation. That information will be put towards a quantitative measurement at each cell. “Adaptive shape segmentation” approach is presented. Adaptive shape segmentation; seeded region growing (SRG) segmentation is a common technique that deals with different shapes in image segmentation. In SRG, the regions grow outwards from the seed points, preferentially, based on the difference between the pixel value and the running mean of values in an adjoining region. This method requires an initial point to be known, which is called the seed. After obtaining the seeds, the process is repeated simultaneously for both foreground and background regions until all the pixels are assigned to either foreground or background. Those pixels that are adjacent to a region are assigned first according to its intensity. “Fig. 1” shows the steps of applying the method.

Figure 1. Flowchart of adaptive shape segmentation

An important advantage of SRG method is that it can be applied to microarray images containing spots of any shape and size.

The seed; which is the initial point to start applying the method is determined automatically. The pixel which has the maximum intensity in the spot window: as that point should most probably one pixel of the inspected spot.

4. Results and Discussions

The presented segmentation method is tested on a number of noisy microarray images drawn from Princeton University Microarray (PUMA) database. Depending on the degree of noise and how the spots are expressed, three noisy microarray images are used. The following figures from “Fig. 2” to “Fig. 4” show results of applying the proposed method on three samples of images having different qualities.

Figure 2. Applying Adaptive Shape Segmentation Automatically to Image 10223Using a Maximum Seed. (Img1: PUMA Experiment ID: 10223, Sub-array 4th row, 2nd column.)

Figure 3. Applying Adaptive Shape Segmentation Automatically to Image 10224 Using a Maximum Seed. (Img2: PUMA Experiment ID: 10224, Sub-array 1st row, 4th column)

Figure 4. Applying Adaptive Shape Segmentation Automatically to Image 10966 Using a Maximum Seed. (Img3: PUMA Experiment ID: 10966, Sub-array 8th row, 4th column)

The accuracy of the proposed method was analyzed by means of a statistical analysis and compared [15] [16] respectively showing that a spot was very efficiently segmented if at least 90% of the entire spot area was enclosed in the contour of that spot. Table 1 shows results of applying the adaptive shape segmentation method using a maximum seed. This way requires an initial point to be known, which is called the seed. An advantage of using SRG in microarray image segmentation is that the location of foreground pixels and background pixels can be estimated. It also deals with different shapes and sizes of spots [17]. By inspecting the results of applying the segmentation method using a maximum seed, it is clear that adaptive shape segmentation using a maximum seed can segment all the three of the gridded microarray images correctly, and give better results with higher accuracy.

Table 1. A Brief Comparison of “Adaptive Circle Segmentation” Results

5. Conclusions and Future Work

The Microarray technology has a great potential in obtaining a deep understanding of the functional organization of cells. The emergence of this technology allows the researchers to tackle difficult problems and reveal promising solutions in many fields. In this respect, our research based heavily on the requirement for a reliable yet time efficient automated method. One of the key steps in extracting information from a microarray image is the segmentation whose aim is to identify which pixels within an image represent which gene. This task is greatly complicated by noise within the image and a wide degree of variation in the values of the pixels belonging to a typical spot. This work emphasizes the impact of microarray image segmentation. In the segmentation stage, efficient segmentation method has been presented; “adaptive shape segmentation” method. When the proposed method inspected, it was found clearly that “adaptive shape segmentation” gives fine results using a maximum seed.