One modified Gabor filter is utilized for carrying out the preprocessing step in this research. The modified discrete wavelet transform is used to feature extraction. Following this, the fitness mutant particle swarm optimization approach is used. The classification is done utilizing the classified namely auto encoder convolutional neural network based on these features. Finally, images affected by a brain tumour are segmented using a modified k means integrated OTSU segmentation method. The following Fig. 1 depicts the overall processing flow of the research study.
3.1. DATASET COLLECTION
The MR brain images used in this study were taken from the BRATS 2015 database. The assessment is based on MR brain images obtained from this database, and the results show that brain tumour segmentation outperforms other current techniques. There are 274 test MR images for experimentation in the BRATS 2015 database. These MR pictures are partitioned into two categories: training and testing. The training data uses 80% of the images in the database, whereas the testing data uses 20% of the images in the database for evaluation.
3.2. PREPROCESSING USING MODIFIED GABOR FILTER
The Gabor filter is used to analyze 2D Xray images, and it is directly connected with the Gabor wavelets. The first and most critical stage in brain tumour prediction is an image enhancement, which uses a database image of a brain tumour as input. There are a lot of rotations and dilations in the Gabor filter, and they take a long time. This results in a blurred image that is unsuitable for subsequent steps of image processing. An MGF technique has been implemented to tackle this flaw. The following are the changes.

•The spatial aspect ratio is not taken at the early stage; the image distortion is reduced.

•The spatial aspect ratio is directly taken into account at the kernel size.

•This assists in obtaining clear images as well as a considerablevariation in PSNR values.
The steps outlined above result in a better overall response. MGF's mathematical model is as follows:
$$\text{f} \left(\text{p}, \text{q}; {\sigma }, {\theta }, {\phi }, {\lambda }\right)=\text{exp}\left[\frac{1}{2}\left(\frac{{\text{c}}_{1}^{2}+{\text{d}}_{1}^{2}}{{{\sigma }}^{2}}\right)\left(\text{cos}\left(2{\pi }\frac{{\text{c}}_{1}}{{\lambda }}\right)+{\phi }\right)+\text{j}\left(\text{s}\text{i}\text{n}\left(2{\pi }\frac{{\text{d}}_{1}}{{\lambda }}\right)+{\phi }\right)\right]$$
where
\({\text{c}}_{1}=2\frac{\text{c}\text{cos}\left({\theta }\right)+\text{d}\text{sin}\left({\theta }\right)}{\text{n}1}\) and
$${\text{d}}_{1}=2\frac{\text{c}\text{s}\text{i}\text{n}\left({\theta }\right)+\text{d}\text{cos}\left({\theta }\right)}{\text{n}1}$$
Where nkernel size
$$\text{M}\left(\text{c},\text{d}\right)= \text{I} \left(\text{c}, \text{d}\right)\text{*} \text{f} (\text{p}, \text{q}, {\sigma }, {\phi }, {\theta }, {\lambda })$$
Where M (c, d) ◊ Modified Gabor filter image
Algorithm
Modified Gabour Filter
1: For brain tumour identification, retrieve an image from a database.
2: The image is reduced by 2 when using the bilinear transformation approach.
3.For the given parameters, find the Gabor filter kernel.
4.In steps of 450, use MGF to input image//theta
5. Add image magnitude for diverse theta value.
6.Bilinear transformation is used to interpolate the sum image by a factor 2.
3.3. MODIFIED DISCRETE WAVELET TRANSFORM BASED FEATURE EXTRACTION
The DWT has the capacity to maintain an image's edge sharpness and can provide local frequency information of an image. This includes two disadvantages like directionality and anisotropy, and so it is difficult to represent images with contours and sharp edges in different directions. Contourlet Transform (CT) is defined to reduce the complexity of the wavelet transform, such that transform should be stable when the input signal shifts or else it should be shiftinvariant.
The NSPFB and the NSDFB structure are the two shiftinvariant components of the NSCT. The pyramid structure is attained using doublechannel 2D (NSFB), for the multiscale property. Low pass filters have been used to create these NSFB. A directional filter bank, which integrates twochannel filter banks and resampling, also provides directionality. The 2D frequency plane is segmented into directional slice using these filters. The shiftinvariant directional expansion is given by the NSDFB. For the formation of NSDFB, upsamplers and downsamplers are removed from the directional filter bank.
A new feature vector is created by utilizing DWT, NSCT and GLCM. The formation of this includes the following steps:
1) Daubechies 1 wavelet at three levels of decomposition is used on image segmentation. The GLCM in four directions (0◦, 45◦, 90◦, and 135◦) is computed using approximate wavelet coefficients App, then evaluate average of these GLCMs (isotropic GLCM) expressed by vector XDWT.
2) The image is segmented into 8 directions with NSDFB at three scales with 'maxflat' pyramidal filter and 'dmaxflat7' directional filter for NSCT. To form a single vector Nsb, all subband coefficients (low pass and 8 directional subband) are added. GLCM in four directions is evaluated from NSCT coefficients Nsb, similar to wavelet coefficients App, and then the average is obtained (isotropic GLCM). Here XNSCT stands for resultant feature vector.
3) The feature vector f = [XDWT, XNSCT] is utilized for the classifiertraining/testing.
3.4. OPTIMAL FEATURE SELECTION USING MUTANT PARTICLE SWARM OPTIMIZATION
The various optimization problems are solved using an intelligencebased random search algorithm known as PSO. The position sets of particle swarm’s are represented as x, and the particle positions are given as x1, x2,…xn. The previously visited positions sets p can be marked as p1, p2,...,pn related to particle position. Mutation breeding is an agriculture procedure that has been extensivelyutilized in new plant variant as 3088 or more mutagenic plant varieties have been released, and it is the method of treating seeds with radiation, TDNA, chemicals,or transposons to create mutants and then choosing the right mutants with desirable characteristics to breed. And this breeding process is suggested to control population diversity and enhance the global optimization capacity of this algorithm
The anticipated MBPSO algorithm is a modification of the classical PSO, in which mutation breeding process is updated p, and pg is substituted with best current particle swarm position at the successive iteration. Assign mutation breeding operation cycle to cm, the number of dimensions to d, maximum optimization iteration number to m, number of particles to n, global best visited position to pg, best visited position setspreviously of each particle to p, and current particle swarm positions to x.
The Genetic Algorithm (GA) and the mutation breeding process are similar in that they both use the mutation process for identifying potential offspring. Nevertheless, there are significant differences between mutation breeding and GA. Crossover operation is not included in the mutation process, and so the computational complexity is lower in comparison to GA. Also, the offspring are chosen based on the technique winnertakeall. As the variation of offspring is easily handled in mutation breeding than in GA, it is suggested to utilize this mutation breeding process for creating p in MBSO
Algorithm
Mutation based Particle Swarm Optimization
Initialize x
Initialize v
p = x
pg = the best position in p
k = 1
while k < = m do
forevery particle, i = 1 To n do
vi = wvi + c1r1(pi − xi) + c2r2(pg − xi)
xi = xi + vi
Evaluate xi
if Mutation breeding operation was performed at last iteration then
pi = xi
else
pi = the best position between xi and pi
end if
end for
if Mutation breeding operation was performed at last iteration then
pg = the best position in p
else
pg = the best position between pg and the best position in p
end if
if Mod(k, cm) = 0 then
Perform mutation breeding operation
For each particle, i = 1 To n do
pi = pg
End for
For every particle, i = 1 To n do
For every dimension, j = 1 To d do
r = random number of 0 To 1
If r < pm then
pij = a random value in the jth dimension range
End if
End for
End for
End if
k = k + 1
end while
3.5. CLASSIFICATION USING AUTO ENCODER CONVOLUTIONAL NEURAL NETWORK
An autoencoder based CNN is used for predicting the tumour in the brain in this approach, and four layers like convolution layer, deconvolution layer, softmax layer and autoencoder layer are employed for attaining low computation overhead and high accuracy in the prediction of brain tumour. Two portions are involved in CNN as in portion 1, pooling and convolution processes are carried out for creating deep features of raw data. In portion 2, the created features are linked to MLP to carry out the classification process. The explanation of these layers are given below
(i) Input layer. It contains N x k neurons, such that k refers to a variable time series and N refers tounivariate series length.
(ii) Convolutional layer.. In this the required operations are performed by convolution filters on specific time of preceding layer. The filter parameters such as filter size k x l, filter numbers m, and convolution stride s, need to be measured on the basis of domain knowledge or simply by experiments. In this layer, a nonlinear transformation function f should be defined. For example, the preceding layer composed of kvariate time series and univariatelength is N, mvariate time series, and univariate lengthis\(⌊\frac{\text{N}\text{l}}{\text{s}}+1⌋\).
(iii) Auto encoder Layer:This is an artificial network for learning efficient data coding, and this performs encoding for data set by avoiding the signal noise. In the reconstructing portion, one representation is formed from the reduced encoding, which is the same as the original input. The autoencoder formis a feedforward, nonRNN that engages in MLP – including an input layer, an output layer. Autoencoders are thus unsupervised learning models (do not need labelled inputs for enabling the learning process).The encoder and the decoder are provided in the autoencoder layer, expressed as transitions Φ and ψ such that:
$${\Psi }:\text{F}\to \text{X}$$
$${\Phi },{\Psi }=\begin{array}{c}\text{arg}\text{m}\text{i}\text{n}\\ {\Phi },{\Psi }\end{array}{‖\text{X}\left({\Psi }\circ {\Phi }\right)\text{X}‖}^{2}$$
In some case, given one hidden layer, encoder stage of an autoencoderconsiders input \(x\in x\in {R}^{d}=X\) and maps it to \(h\in {R}^{p}=F\) :
(iv) Softmax layer: This layer contains same amount of nodes as output layer. A neural network can be able to generate a probability that a dog is in the image or not, but it would do so for each input separately. Also, this network can run a multiclass function and determine the probability that the dog and also other objects present in the image. These layers are excellent at calculating multiclass probabilities, but they have limitations. As the number of classes grows, this can become expensive, and so Candidate sampling can be performed. The availability of its calculations to a specific set of classes is minimized by using candidate sampling. When determining whether an image of a bowl of fruit contains apples, for example, the probability does not need to be calculated for each type of fruit, and only the apples need to be calculated. Furthermore, each class is provided with one member, so it will not work in situations where an object belongs to multiple classes. In that case, using multiple logistic regressions as an alternative is a viable option.
(v) Output layer.It contains n neurons related to n time series classes, and is entirely linked to feature layer. The maximum output neuron is chosen as the input class label emotion in classification process.
Training of auto encoder CNN
The CNN is trained through training samples sequence \(\left(\right({x}_{1}, {y}_{1}), ({x}_{2}, {y}_{2}), . . . , ({x}_{N}, {y}_{N}\left)\right)\)with\({x}_{t}\in {R}^{N\times k}\), \({y}_{t}\in {R}^{n}\)for \(1\le t\le N\). The network is trained with several steps:
Step 1Network establishment. As shown in Fig. 1, identify the CNN architecture consists of two convolutional layers and two pooling layers. Assign the number of neurons in the input and output layers based on classification task. Set CNN parameters as default values. Use a small random number to start the weights and bias. Choose a learning rate η and an activation function f, with the sigmoid function as a common example:
$$f\left(x\right)=sigmoid\left(x\right)=\frac{1}{1+{e}^{x}}$$
11
Step 2 Randomly select training sample from training set.
Step 3 Computelayer output

The Convolutional layer outputis given as
$${C}_{r}\left(t\right)=f\left(\sum _{i=1}^{l}\sum _{j=1}^{k}x\left(i+s\left(t1\right),j\right){\omega }_{r}\left(i,j\right)+b\left(r\right)\right)$$
12
here\(x\in {R}^{N\times k}\)shows the input higherorder semantic features or preceding layer output, s represents the convolution stride, \({C}_{r}\left(t\right)=\)tth component of rth feature map, ωr∈Rl×kand b(r) denote weights and bias of rth convolution filter.

The pooling layer output isprovided as
$${P}_{r}\left(t\right)=fg\left({C}_{r}(\left(t1\right)l+1\right),{C}_{r}\left(\left(t1\right)l+2\right),\dots ,{C}_{r}\left(tl\right))$$
13
In which the function g shows the pooling strategy, the most popular used is averaging or max pooling. Here the pooling process reduces the point data, without altering the number of feature maps.

The final layer output isprovided as
$$O\left(j\right)=f\left(\sum _{i=1}^{M}z\left(i\right){\omega }_{f}\left(i,j\right)+{b}_{f}\left(j\right)\right), j=\text{1,2},\dots ,n$$
14
In whichz refers the final feature map in the feature layer, \({b}_{f}\)shows output layer bias and \({\omega }_{f} \in {R}^{M\times n}\)denotes the connection weights among feature and output layer. Therefore,MSE is given as
$$E=\frac{1}{2}\sum _{k=1}^{n}e{\left(k\right)}^{2}=\frac{1}{2}\sum _{k=1}^{n}(O{\left(k\right)y\left(k\right))}^{2}$$
15
Step 4 Update bias and weightwith gradient descent method.
$$\text{p}=\text{p}{\eta }\frac{\partial \text{E}}{\partial \text{p}}$$
16
In which p represents parameter value, and p denotes to ω_r,ω_f ,b, or b_f in this CNN.
Step 5select successive training samples and go to Step 3 until samples
aretrained.
Step 6Increase the number of iterations. If the iteration number is equivalent to predefined maximum value, stop the algorithm. Else, go to Step 2. Based on the above steps, social emotion is classified.
3.6. SEGMENTATION USING K MEANS INTEGRATED OTSU ALGORITHM
As people are interested in specific parts of an image as segmentation is an essential method in image processing applications. It separates an image into distinct areas with high similarity between pixels in each area and high contrast between areas. The kmeans integrated OTSU segmentation algorithm is used for brain tumour segmentation in this research.
Algorithm
K means integrated OTSU algorithm
Step 1: Assign input asbrain tumour image. The inputted image is converted to HSV color space.
Step 2: Extract V channel of HSV color space.
Step 3:Initialize Separation Factor (SF) = 0 and N = 2, in which N denotes the number of classes. The SF value lies among 0 and 1. The largest SF valuedeterminesthe imageasaccurately segmented.
Step 4:Use Otsu's thresholding on the V channel of the HSV colour space. Ostu devised maximum class variance approach to simplify computation, minimise time, and improve the process's effectiveness. The image is segmented by choosing the threshold valueautomaticallywith highest interclass variation among the target and background and lowering class variance\({{\sigma }}_{\text{W}}^{2}\), and here the total variance (\({{\sigma }}_{\text{B}}^{2}+{{\sigma }}_{\text{w}}^{2}={{\sigma }}^{2}\)) ), i.e. the sum of withinclass and betweenclass variance are constant for the specific image.
Step 5:The value of SF, which is defined as\({{\sigma }}_{\text{B}}^{2}/{{\sigma }}^{2}\) is identified. If value SF is 1, then image is segmented, else, number of classes is incremented by 1 i.e. N = N + 1, and Otsu's thresholding is usedin the N value.
Step 6:The Otsu's thresholdingoutput shows oversegmentation. Therefore, certainapproaches aremerged over the segmented regions. Here, Kmeans clusteringis depicted asa partitioning approach for clusteringobjects, this withingroup variance is reduced. The steps are provided below:
a.Randomly assign twoclass centres to specify group centroids.
b.Compute histogram bin value distance amongevery image pixel and class centroids; allocate image pixel to class centroid nearest to it.
c.Calculate the mean histogram bin value of the same group to recalculate the new centroids positions.
d. Steps b and c should be repeated until the value of the centroids changes