Image Segmentation

Medical image assessment based on specific computer algorithms is gaining popularity due to its accuracy and adaptability to a variety of images with different dimensions and orientations. During image assessment, the extraction and evaluation of the infected section using a suitable procedure is largely implemented to examine the greyscale and RGB-scale images [1-4].

Segmentation is one of the important image processing techniques implemented to extract a particular section of the image. A considerable number of the SOI extraction techniques are available in the literature. Based on the implementation, it can be classified as (i) automated segmentation and (ii) semi-automated segmentation techniques, the choice of which depends on the expertise of the operator and the complexity of the test image to be examined [5-7].

Requirement of Image Segmentation

In medical image assessment, the segmentation implemented is normally used to extract the infected section from the clinical-grade test image for the evaluation and treatment planning process. The combination of multi-level thresholding and segmentation is executed by the researchers to examine a class of diseases using the images. These processes also help form a hybrid image processing procedure which offers better evaluation compared to traditional techniques. The features extracted from the segmented image section are considered to develop and implement a class of Machine Learning (ML) systems, the accuracy of which depends mainly on the segmented disease section. Further, the features of the segmented section can be considered to improve disease detection accuracy in the machine learning system [8-10].

The major requirements of image segmentation are listed below:

i. Extraction and assessment of the disease-infected section from the two- dimensional clinical images. Figure 4.1 depicts the hybrid image processing scheme implemented to extract and evaluate the skin melanoma section from the digital dermoscopy image. The task is to extract the melanoma section which is achieved using a threshold process to enhance the image and to extract the SOI with a segmentation procedure. After extracting the essential section from the image, its value is then compared to the ground-truth and, based on the attained quality values, the performance of the segmentation technique is validated.

ii. Implementing and improving disease detection accuracy of the ML system. During this stage, the image features are initially extracted with a technique

Hybrid Image Processing Technique Implemented to Examine the Skin Melanoma

FIGURE 4.1 Hybrid Image Processing Technique Implemented to Examine the Skin Melanoma.

and, based on these features, classification systems are implemented and trained to support the automated detection of the disease,

iii. Improving the performance of the traditional and modern DL system using the image features obtained from the segmented disease section.

Extraction of Image Regions with Segmentation

Extraction of the Section of Interest (SOI) from the medical image is the essential process which helps assess the disease in images with better diagnostic accuracy irrespective of the imaging modality. To achieve this, several segmentation procedures are proposed and implemented in the literature. This section presents a few commonly used image segmentation procedures employed in medical image processing.

Morphological Approach

Morphological approach is used to pre-process images considered for the investigation through several operations implemented on the image pixels. This technique is widely utilized to smoothen the surface of the image. This procedure is a commonly used segmentation process, adopted to extract the particular pixel group based on what is needed.

Morphological enhancement and segmentation is one of the oldest image processing schemes. Due to its merit, morphology-based segmentation techniques are widely employed in a variety of segmentation procedures, one of which is the Markov Random Field segmentation (MRF) with Expectation Maximization (EM). The MRF-EM is an image segmentation technique employed to extract image regions into various sections. Information regarding the MRF-EM-based segmentation can be found in [11,12].

MRF-EM is a generally considered method for greyscale image segmentation problems. The necessary details are presented below:

Consider a greyscale trial picture /= {Y(M. N) й Y й L - in which Y symbolizes the strength of the picture at the pixel location (M, N) and L denotes the number of thresholds of the image (normally 256). Through segmentation, MRF will estimate the arrangement of every pixel by mapping them into a cluster of

arbitrary labels defined as X = {xi.....ед)|л, € I. The number of labels is chosen as

three in this work, which separates the test image into three sections. Implementation of MRF-EM algorithm is defined below:

Step 1: Set the amount of labels (/) based on quantity of threshold values (7) Step 2: Arrangement of cluster classes based on the selected /

Step 3: Determine the initial parameter set 0|O)l' and likelihood probability function p(<>> (/Jb'i).

Step 4: Update the MRF model x<0 such that, the energy function U is minimised.

where, N, is a four pixel neighborhood and Vc is the clique potential.

Step 5: Implement the EM operation to update the parameter set 0(,) constantly till the log likelihood of p(n (f[x) is maximized.

Step 6: Display the separatee! labels as the segmented results.

The performance of the MRF-EM algorithm is initially tested on greyscale and RGB-scale test images and then it is implemented on the images associated with the Salt & Pepper noise. Figure 4.2 depicts the brain MRI slice considered for the assessment and the aim is to segment the tumour section from the test image. This work employs a hybrid technique which integrates the morphology-based image enhancement using the Expectation Maximization (EM) operation and then employs a segmentation procedure, which separates the image section into three parts: normal brain section, tumor, and the background. The morphology and the EM help enhance the test image by executing a suitable morphological operation. This helps group the pixels to enhance the SOI to be extracted and evaluated.

Figure 4.2 (a) presents the test image. The MRF-EM segmentation is employed by fixing the iteration level at three. Initially, the EM algorithm is executed to improve the image and it creates the essential image labels (processed image) based on the EM value. The convergence of the EM operation is depicted in Figure 4.2 (b) and the enhanced image is presented in Figure 4.2 (c).

Segmentation Results Attained for the Brain MRI Slice

FIGURE 4.2 Segmentation Results Attained for the Brain MRI Slice.

The segmentation results attained with the MRF-EM are illustrated in Figure 4.2 (d) to Figure 4.2 (f), where section 1 presents the normal brain section, section 2 depicts the tumor (SOI), and section 3 presents the background. After determining the tumor section, it is examined by a doctor to evaluate the severity of the disease.

The image is then tested against the medical image with the noise. The result is shown in Figure 4.3 (a) to (f). In this image, Figure 4.3 (b) is the initial morphological operation along with the noise. This operation is repeated to eliminate the noise pixels from the image to get the final enhancement image as in Figure 4.3 (c). The remaining results are similar to Figure 4.2. This confirms that the proposed procedure works well on the test image with/without noise, extracting the SOI accurately. The MRF- EM segmentation is also tested on the RGB-scale image (fundus retinal image), the results of which are depicted in Figure 4.4. This result also confirms that the MRF- EM approach works well on a class of test images, helping achieve better segmentation results to improve the disease diagnosis process. More information on the MRF-EM can be found in Rajinikanth et al. [13]. It is an automated segmentation technique existing to solve medical image analysis problems.

Circle Detection

Plough Transform (HT) is a procedure meant to recognize the circles/circular- shaped objects from digital images [14-16].

Segmentation Results Attained from the Brain MRI Slice Associated with Noise

FIGURE 4.3 Segmentation Results Attained from the Brain MRI Slice Associated with Noise.

The HT employed in this work is discussed in [17].

Let, r = radius, X and Y = illustration axis, A and В = centre of a random circle. Then, it can be expressed as

The HT is able to recognize the X and Y based on the preferred A, В, and r. The HT locates and traces the binary pixels (Is) in the test image after a possible border discovery procedure. The conventional HT traces and extracts all the existing extremely noticeable binary pixels (Is) in the digital illustration. In most cases, this HT fails to mine leukocyte segments from the hematological picture. HT-based circle detection procedure can be found in [18,19]. To increase the success rate in HT-based circle detection, the morphological procedure is included to eradicate the unnecessary but noticeable pixels in the pre-processed image.

The following steps present the modifications executed in HT to improve its segmentation accuracy:

Step 1: Consider the pre-processed hematological illustration and execute border discovery.

Step 2: Implement the HT with a selected circle radius and recognize the entire pixel group with a magnitude unity.

Step 3: Extract the Haralick features and recognize the major-axis from the identified pixel groups.

Step 4: Implement the morphological action to eradicate the pixel groups, which is less than the identified major-axis.

Step 5: Improve the leukocyte segment with a pixel-level comparison.

Step 6: Extract the existing pixel collection and validate with the traces made by the HT.

To demonstrate the HT-based circle detection process, the RGB-scale image depicted in Figure 4.5 (a) is considered. This image shows the availability of the leukocyte along with the blood stains. The primary objective is to extract the leukocyte with considerable accuracy. First, the test image is improved using a thresholding process, which will helps improve the visibility of the test image as depicted in Figure 4.5 (b). This step improves the stained leukocyte segment of the image considerably. Second, the segmentation based on the HT is executed to extract the leukocyte region as in Figure 4.5 (c) and (d). This step consists of procedures such as (i) detection of all the likely sections using HT. (ii) assessment of the section with large major-axis by means of the Haralick algorithm, (iii) implementing the morphological dilation and erosion of the pixel groups whose aspect is smaller than the major axis, and (iv) discovering and mining the binary form of the leukocyte. After extracting the leukocyte, the performance of the implemented HT is validated using a combative analysis between the extracted section and the ground-truth provided by an expert. Essential information on the HT-based circle detection can be found in [16-19].

Detection and Extraction of Leukocyte Using HT

FIGURE 4.5 Detection and Extraction of Leukocyte Using HT.

< Prev   CONTENTS   Source   Next >