The Proposed GLCM-PSO-SVM Model

The block diagram of the developed GLCM-PSO-SVM system is applicable in predicting tumors in MRI as depicted in Fig. 10.1. In this model, an individual is forecasted completely by exploiting medical tools wdth IoT gadgets. After capturing the MRI brain images, it can be preprocessed. Hence, the features were extracted from the preprocessed image. Class labeling is more applicable to train the model. If a method is trained exactly, the testing process is initialized w'hen images are classified in corresponding labels. The trained model is capable of testing the provided MRI brain images effectively.


Brain MRI is subjected to corruption by noise w'hile processing image transmission as well as digitization as shown in Fig. 10.2. Preprocessing is operated to eliminate the noises from MRI. The extracranial tissues such as bone, skin, and hair are rejected and result in the conversion of heterogeneous images to homogeneous images. Noise can be eliminated wdth the application of filters and even corrupted

Working process of the GLCM-PSO-SVM model

FIGURE 10.1 Working process of the GLCM-PSO-SVM model.

with minimum data of an image. Additionally, traditional filters make a smoother image constantly and develop stronger edges of an image.

It is applied with an anisotropic diffusion filter to preprocess the brain MRI since the filter avoids the noise and saves the edges. In the case of a corrupted image, the features are blurred. Here, it is employed with anisotropic diffusion filtering for denoising. The filter ranges the neighboring pixels on the basis of intensity, and a median value can be computed for pixels after estimation. A novel median value has been replaced using a central pixel.

Feature Extraction

The GLCM is a model matrix utilized to determine the texture design with modeling texture as the 2D array gray-level difference. These attributes exist between a

Preprocessing output (a) input image, (b) preprocessed image

FIGURE 10.2 Preprocessing output (a) input image, (b) preprocessed image.

group of features to determine the pixel contrast as well as the power of ROI. This feature is more useful for separating the normal tissue from anomalous tissue based on removed contrast and power. It is assumed to be the statistical approach that considers spatial connection between the pixels, because it is a gray-level spatial dependence matrix. The GLCM facet is computed in four directions: 0°, 45°, 90°, and 135°; and for four distances: 1, 2, 3, and 4. Also, there are four characteristics in the GLCM (contrast, energy, homogeneity, and correlation) and can be estimated with the given functions

The number of gray levels used in computing is the GLCM size. A matrix unit P(i,j I Av, Ay) is said to be a relative frequency, and somewhere two pixels are separated with pixel distance (Av. Ay) in the provided region, with intensity i and intensity j.

The GLCM P[i,j] is comprised of a location of pixels by related gray level and referred to as identifying displacement vector d = (dx, dy) and calculates each pair of pixels that has been partitioned using d that is comprised of gray levels of / and j. Wavelet is determined as a mathematical function signifying tiny waves that are scaled, as well as shifted models called a mother wavelet

where a implies a scaling variable and b denotes as a shifting parameter. The wavelet transform (WT) considers the image on different resolution scales and changes the image to a multiresolution image through various frequency modules. The wavelet is said to be discontinuous and refers to a step function. For a function /, the Haar WT is determined as

where L represents the disintegration level, a signifies the approximation sub-band, and cl mimics the extended sub-band. The WT has been implemented to each row and column of the image accomplished from the previous level. The final has been classified as four sub-bands: LL, HL. LH, and, where L = Low, H = High. An estimation of actual image is added in the LL sub-band, while alternate sub-bands are constrained with missing rates. The LL sub-band can be attained at diverse stages that have been degraded as LL, HL, LH, and HH sub-bands. Law’s energy texture features states that initially, ID kernels of brain image are reformed to 2D filter kernels. The second step has filtering of input mammogram images under the application of Law’s 2D kernels and processes the energy of an image.

PSO-SVM-Based Classification

SVM Classifier

The SVM is constrained with maximum accuracy, simple mathematical tractabil- ity, geometrical representation, and so on. Thus, solutions attained are unique and global to eliminate the convergence of local minimum executed under the application of alternate types of statistical learning methods such as NN. It is offered with a /^-dimensional training data set of N size with respect to

where y„ is a - 1 or 1 by means of class 1 or 2. Each ,v„ is said to be a /^-dimensional vector. A maximum-margin hyperplane, which segments class 1 from class 2, is highly applicable in SVM. Hence, a type of hyperplane is described as

where w implies a normal vector of hyperplane and b signifies the bias. It highly concentrates on selecting tv and b to improve the margin among two parallel hyperplanes of higher size while isolating the data. Therefore, two parallel hyperplanes are expressed using the given equation

This model could be converted into an optimizing problem. The purpose of this technique is to boost the distance of two parallel hyperplanes and remove the data falling from the margin. Under the application of elegant mathematical data, this issue can be resulted in

Practically, || vv || is replaced by applying

The major reason depends upon the function of || w ||, which is relevant in calculating the square root value. Once the value has been computed, the solution cannot be changed, and modify the problems as optimizing quadratic programming, which becomes simpler by applying Lagrange multipliers and reputed quadratic programming techniques. It undergoes mapping of binary classification issues, expressed as

where C is a regularizing variable and represents the penalizing relaxation parameter.

Parameter Optimization of SVM Using the PSO Algorithm

For validating the optimal parameter of C, the trial-and-error method is implemented. Therefore, implementation of these techniques tends to overhead in action without any assurance of attaining optimized outcomes. It is the application of PSO that is employed for uniquely optimizing parameters. It is linked by global optimization, accomplished from learning fish training or bird flocking. Then, the cross- validation (CV) model is deployed to invent the fitness function (FF) that is applied in PSO. It encloses a swarm of particles that is extended for all iterations. To attain an optimized result, every particle updates its best position (phcsl) and best global position in swarm (ghesl) as defined below

where / is the particle index, P implies the overall particles, к denotes the iteration index, z shows the recent iteration value, and p represents the position. The velocity as well as position of particles is improved using the provided functions

where v is velocity. The inertia weight vv is applied to manage the global searching as well as a local application. The r, and r2 are distributed in a uniform model by random variables in a range (0, 1). The c, and c2 are positive constant variables named as acceleration coefficients. Thus, the particle undergoes encoding as embedded by parameters C.

< Prev   CONTENTS   Source   Next >