A Novel Approach for Hybrid Image Segmentation GCPSO: FCM Techniques for MRI Brain Tumour Identification and Classification

In recent times, the early detection of brain tumour analysis and classification has become a very vital part of the medical field. The MRI scan image is the most significant tool to study brain tissue for proper diagnosis and efficient treatment planning to detect the early stages. In this research study, the two contributions were executed in the preprocessing mode. (a) Using wavelet transform to apply decomposed sub-bands of a low-frequency signal to control and adapt the spatial and intensity parameters in a bilateral filter and (b) to detect texture regions and block boundary to control and adapt the spatial and intensity parameters in a bilateral filter When compared to other image resolution methods, the adaptive bilateral method restores the original image quality and has a higher accuracy rate. Using the hybrid segmentation method of GCPSO (Guaranteed Convergence Particle Swarm Optimization) -FCM (Fuzzy C-Mean) techniques, the results were compared with various segmentation. The proposed segmentation gives a better accuracy rate of 95.32%.


Introduction
Early detection and treatment of these illnesses is extremely helpful to fnd out the abnormal cells in that image. Te aim of the research process, MRI scan images were used to fgure out that a person had a brain tumour. Te MR images can be used to measure the size of the tumour cells. Tere are a number of ways to use segmentation to fnd brain tumour cells in an MRI scan image. On dynamic contrast-enhanced (DCE) MRI of the breast [1], a semiautomated and interactive technique based on the spatial Fuzzy C-Means (sFCM) algorithm is suggested [2]. Te suggested method is used to obtain a segmented tumour region that is visible to the doctor and provides them with more information about the tumour for their diagnosis. Te suggested method makes use of image fltering, threshold-based segmentation, morphological operations, and pixel subtraction. Te suggested method is predicated on getting distinct images of the skull, brain, and tumour. When compared, the suggested strategy produced a better outcome than the alternative strategy.
Te adaptive threshold-based technique, an alternative to a fxed threshold, can lessen these dependencies by considering the tumour to [3] background ratio (TBR) and tumour size. Because the parameters for the adaptive methods need to be calibrated for each PETcamera system, it is difcult to employ the same methodology on diferent PET systems to get comparable results (e.g., scanner geometry, image acquisition protocol, and reconstruction methodology). Reports indicate that the performance of adaptive techniques underperforms for smaller volumes with lower TBR and SNR. Based on an improved direct region identifcation method, to [4] address MRgFUS post-treatment segmentation challenges for fbroid segmentation in MR images. In the recommended incremental process, an adaptive region-expanding procedure uses the split-and-merge method outcomes as various seed-region selections. Te recommended technique isolates many fbroids with diferent pixel intensities, even within the same MR image. Utilizing area-and distance-based measures, the strategy was evaluated and compared to other studies of a similar kind in the literature.
Otsu thresholding, Discrete Wavelet Transform (DWT), Principal Component Analysis (PCA), Support Vector Machines (SVMs), Least Squared-Support Vector Machines (LS-SVMs), Proximal Support Vector Machines (PSVMs), and Twin Support Vector Machines (TWSVMs) are employed in the segmentation phase, feature extraction phase, feature reduction phase, and classifcation phase, respectively. Te topic of creating active contour models and segmentation algorithms for automatically spotting brain cancers in multimodal volumetric MRI data. Using active contour and region-growing segmentation, we improved the detection of numerous brain tumours.
Comparing the suggested quantum-inspired classifer to traditional machine learning classifers, it is a novel and superior methodology for evaluating clonogenic experiments [5]. It is shown to be an important feature in detecting difcult cell colonies among the extracted and analyzed image features. Te 11 layers of the SegCNN model were developed using the Adam optimization solver and a predetermined batch size. Te proposed model delivers a 98% dice similarity coefcient (DSC), which is noticeably superior to the most recent research reports. In the present research work, a diagnosis of brain tumour cell detection using several methods has been implemented for image denoising and image segmenting. For this research, the past ten years of research work were reviewed and discussed for comparison purposes. Te proposed techniques are implemented and get more accurate results when compared with previous work. Tis work provides help to diagnose diseases in a very small period of time and with better accuracy and value, so this study will contribute more efectively in the ground of image processing. Te research methodology of this work is as follows: (1) Te image database is generated by the collection of cancerous images from a private cancer hospital. 130 images are collected from private centers. (2) To be developed a noise reduction algorithm using an Adaptive Bilateral Multiresolution technique that is more efcient in improving image resolution.
(1) A Adaptive Bilateral Multiresolution method is combined with a wavelet transform threshold. (2) To eliminate compression artifacts using spatial adaptive methods and to avoid over-smoothing of original image texture regions while efectively removing ringing and blocking artifacts.
(3) Using a hybrid segmentation technique to segment a fltered image using GCPSO (Guaranteed Convergence Particle Swarm Optimization) and FCM (Fuzzy C-Mean) techniques was compared with various segmentations. Te proposed segmentation gives a better accuracy rate of 95%.
Te motivation of this study is discussed with (a) the uncertainty and impreciseness in detecting tumour formation, orientation, and location; (b) the use of the "Divide and conquer" strategy on MRI imaging to perform segmentation; (c) access to analysis and easily deployable tools such as MATLAB packages; (d) lacks powerful feature extraction through multiple orientations and efcient classifcation. Te proposed work is shown in Figure 1.

Related Works
Biratu et al. has suggested utilising intersection over union, similarity score, and accuracy against the real world to evaluate ROI [6]. Te ROIs were compared with several deep learning and region-based techniques. Tey generated a DSS value by randomly choosing an image to test. Te seed point initiation and obtaining a better ROI for any photos of a brain tumour provide the key difculties of this work of region developing segmentation method. Te seed point was created to automatically produce for the major approaches of this study's random image selection. Shahvaran et al. has used multimodal tumour pictures to perform automatic [7] tumour extraction. Using the active contour approach, which is utilised for tumour disease extraction based on the border of the discovered brain tumour regions, it is possible to detect the tumour image using the K-means clustering algorithm. On the BraTS 2013 dataset, which included patients with both high-grade and low-grade tumours, the suggested technique was assessed. Te suggested method outperformed existing active contour based methods when it came to segmenting tumours, with mean Dice similarity coefcients for high-grade and low-grade tumours of 0.9179 (0.025) and 0.8910 (0.042), respectively. Arun Kumar et al. [8] developed an outstanding performance in brain tumour classifcation and found the best accuracy rate. Te classifcation techniques are built on the classic machine vision approaches, including image enhancement, Fourier transform, segmentation, HOG, ANN classifcation, and feature extraction. So, the classifcation method gives a 92.14% accuracy using K-fold cross validation techniques.
Breast cancer is a dreadful disease. Early detection of ductal carcinoma (breast cancer) will save lives. Tis is possible by amalgamating microwaves and infrared thermography using a Convolutional Neural Network (CNN). Microwaves and infrared thermography are used as [9,10] sources of radiation and heat imaging recorders, respectively. By transmitting the radiant energy to the mammary gland with the electrical properties of normal and abnormal tissue, the magnetic permeability of normal tissue contradicts with the ill one, which has a radius of 5 mm, and its location can be found by placing a sensitive screen below the mamma. Hence, the heat distribution pattern can be traced out. Tis technique is enhanced by implementing a Machine Learning (ML) model. Terefore, even a slight variation in heat pattern can be detected.
Usually, BTs (brain tumours) are categorized using MRI (Magnetic Resonance Imaging). In this paper, two databanks are used in which Gaussian Convolutional Neural Network (GCNN) is applied [11]. Te frst dataset is used to group tumours into glioma, meningioma, and pituitary. With the second database, glioma is disparate into three grades. Te frst and second databases consist of 3064 and 516 images, correspondingly, that are elicited from T1-weighted complexity-improved images. Te datasets are interconnected after performing GCNN. Using data augmentation, an accuracy of 97.14% in the frst and 99.8% in the second databank is achieved.

Adaptive Bilateral of Multi-Resolution Framework.
A multiresolution analysis is the best tool for reducing noise in signals. So, it is more possible to diferentiate image information, and noise is restored at one resolution level than another. As proposed, the adaptive bilateral method of the multiresolution framework is applied which decomposes a signal into frequency sub-bands with wavelet decomposition. A signal is back reconstructed; a bilateral method is applied to the sub-bands and the adaptive bilateral method of multiresolution was reducing low-frequency noise components. Applying wavelet threshold to the entire subbands, it is reducing the image noise very efectively in the combination of adaptive bilateral multi-resolution and wavelet threshold.

Compression Blocking Artifacts Reduction.
Te two parameters of the spatial domain (σ d ) and intensity domain (σ d ) are the control and behavior of the standard bilateral technique. In this case, the parameters are chosen carefully for the compression artifact reduction.

Analysis of Parameter.
Te two parameters of the spatial domain (σ d ) and intensity domain (σ d ) are control and behaviour of the standard bilateral flter. In the case of a bilateral flter, parameters are chosen carefully for the compression artifact reduction.
Te edge signal shows the frst subplot; the edge discontinuity value is 10. Te output of the intensity domain σ r values shown in the second subplot in that fgure. If the intensity domain σ r value is less than the discontinuity, basically the flter is useless to reducing the discontinuity. If the intensity domain σ r value is larger than the discontinuity, it reducing the discontinuity.
Te r value in the spatial domain can be utilised to control and enhance smoothness. If the d value of the spatial domain is greater than the level of discontinuity, the smoothing will be more extensive. It is possible to lessen the discontinuity without using d's solution if the spatial discontinuity is less than the value of d. As a result, we take measurements of the block boundaries, r's value, and the degree of discontinuity. Te texture region would be too smoothed if the d value were small; instead, it is huge. Te nonadaptive bilateral flter will result in numerous issues, such as too blurred texture regions when a strong parameter is used to reduce the blockiest. Te blocking artefacts are not there when we choose a lesser setting.
Te control and behavior of the standard bilateral flter are controlled by the two parameters of the spatial domain (d) and the intensity domain (d). In the case of bilateral flter parameters are chosen carefully for compression artifact reduction.
Te edge signal shows the frst subplot; the edge discontinuity value is 10. Te output of intensity domain r values is shown in the second subplot in that fgure. If the intensity domain r value is less than the discontinuity, the flter is useless to reduce the discontinuity. If the intensity domain r value is larger than the discontinuity, it will reduce the discontinuity.
At that same time, enhanced smoothing can be controlled by the spatial domain r value. Te spatial domain d value is larger than the discontinuity, the wider the extent of the smoothing. If the spatial domain d value is less than the Computational Intelligence and Neuroscience discontinuity, then reducing the discontinuity has no solution. As a result, we measure the discontinuity amount along with the block boundaries and the value of r accordingly. To avoid over-smoothing the texture region, the d value is large; otherwise, it is small. Te nonadaptive bilateral flter will create some issues, like when we choose strong parameters to reduce the blockiest; the texture region is over-blurred. When we choose weaker parameters, it does not completely remove the blocking artifacts.
Here, Figure 2 shows the parameter values of σ d and σ d on block discontinuity are illustrated. Te input signal value is 10. Te performance of the middle image shows the parameter σ d value and diferent σ r value. Te performance of the last image shows σ r value and diferent σ d value. Previously we discussed, the two modules are included in this framework. Te frst module is detecting the discontinuity amount and adjusts the parameter σ r value. Te second module is detecting the smoothness of a texture regions and adjust the parameter σ d value. Te new framework of the adaptive bilateral method in Figure 3.
Here, Figure 2 shows the parameter values of d and d on block discontinuity are illustrated. Te input signal value is 10. Te performance of the middle image shows the parameter d value and diferent r value. Te performance of the last image shows r value and diferent d value. Previously we discussed, the two modules are included in this framework. Te frst module is detecting the discontinuity amount and adjusts the parameter r value. Te second module is detecting the smoothness of a texture regions and adjust the parameter d value. Te new framework of the adaptive bilateral method in Figure 3.
To detect block, discontinuities, the vertical boundaries of an input image is fltered with the value of [−1, 0, 1] and the horizontal boundaries with [−1, 0, 1] T the value of the results. Te parameter σ r should be at least equal to efective values. Te block discontinuities are detected in the boundaries: if the bilateral flter is applied to the boundary only, the block cannot reduce. If we reduced the block very efectively, entire blocks of the bilateral flter should be applied.

Hybrid
Segmentation. An enhanced EDPSO (Darwinian particle swarm optimization) and Quantum Entanglement and Wormhole Behaved Particle Swarm Optimization Techniques [12][13][14] were proposed to segment a tumour image which overcomes this existing method of GCPSO (Guaranteed Convergence Particle Swarm Optimization). We propose, a new Hybrid GCPSO (Guaranteed Convergence Particle Swarm Optimization)-FCM (Fuzzy C-Mean) algorithm use to each particle to every number of generations/iterations so the ftness value of all the particles to improve.
Hybrid GCPSO algorithm for fuzzy clustering is as follows: (1) Initializing the attributes of GCPSO and Fuzzy c-mean (2) Initializing the attributes of population P, c1, c2, m, and w (3) Generate particle attribute represent P (x, gbest, p best, and velocity are n * c matrices)

GCPSO
(i) Compute the cluster center for the whole particles (ii) Compute the ftness value of all the particles (iii) Calculate pbest for all the particles (iv) Evaluate gbestfor the swarm (v) Entire particle velocity to be evaluated (vi) Entire particle position to be evaluated

Fuzzy C-Mean
(i) Compute the center of the cluster for the each particles (ii) Find the Euclidian Distance method to represent the attribute d ij , the i varies from 1, 2, . . ., n; and the j varies from 1, 2, . . ., c; for each particle using the following equation: (iii) Evaluate of −++ the function µ ij , i varies from 1, 2, . . ., n; and the j varies from 1, 2, . . ., c; for the entire particle.
(iv) All the particle of pbest to be evaluated (v) Swarm of gbest to be evaluated   Te GCPSO (Guaranteed Convergence Particle Swarm Optimization)-FCM (Fuzzy C-Mean) algorithm focuses on a new particle deal with a present best position in the neighbourhood. In the present research study, each element considers as a member of the swarm, the following a new particle for the velocity update equation is given as follows: To improve the random search area in the region of the best position was improved. Te random search area of the random vector and random diameter r and ρ(t).
Here, the term # sc is a success and # fc is failure is defned as a number of consecutive failure and success. Since high dimensional search spaces in a few iteration sc values are 15 and fc value is 5.

Enhanced Fuzzy C-Mean Techniques. FCM is clustering methods which allows for partitioning into multiple clusters.
Te minimization of the objective function: where m � actual number higher than the value of 1. u ij � x i is a degree of partisanship in cluster j, Now the x i is the number of i th d-dimensional measured data. c j � Cluster center. Enhanced Fuzzy C-Mean technique is used to calculate the middle of the cluster for entire particles and the Euclidian Distance method d ij , the i values are 1, 2, . . ., n; and the j values are 1, 2, . . ., c; the following equation is To upgrade the function µ ij , i values are 1, 2, . . ., n; and the j values are 1, 2, . . ., c; for the entire particle. Te following equation is Finally, evaluate the best position (pbest) for entire particles and evaluate the global best position (gbest) for the swarm. If Fuzzy C-Mean Clustering is the termination state GOTO initial stage of the Fuzzy C-Mean Clustering and If the GCPSO terminating state is not met, GOTO initial stage of the method. In the current research, 130 diferent brain tumour test photos are analyzed. Utilizing the suggested approach of multiresolution of the adaptive bilateral technique, reduce noise and improve image quality. In this research work, the proposed hybrid segmentation approach of Guaranteed Convergence Particle Swarm Optimization (BMPSO) and Fuzzy C-Mean (FCM) was compared and found to be the more efective method. Tis new framework, a hybrid BMPSO-FCM, has assessed an accuracy rate of 95.92 percent for detecting brain tumour cells.
Finally, evaluate the optimal particle position (pbest) and the optimal global particle position (gbest) for the swarm. If Fuzzy C-Mean Clustering is in the termination state, proceed to the initial stage of Fuzzy C-Mean Clustering. Similarly, if Guaranteed Convergence Particle Swarm Optimization is in the termination state, proceed to the initial stage of the procedure. In the current study, 130 distinct test images of brain tumours are analyzed. Reduce noise and enhance image quality by implementing the suggested multi-resolution adaptive bilateral technique. In this study, the hybrid segmentation methodology of Guaranteed Convergence Particle Swarm Optimization (BMPSO) and Fuzzy C-Mean (FCM) was determined to be the superior method. Tis new framework, a hybrid BMPSO-FCM, ofers a 95.92 percent detection accuracy rate for brain tumour cells.

Experimental Result
A multiresolution analysis is the most important tool for reducing noise in signals. It is possible to diferentiate image information and noise is better than one resolution level than another. As proposed, the adaptive bilateral method of the multiresolution framework is applied it is shown in Figure4, decomposed a signal into frequency sub-bands with wavelet decomposition. Based on these trials, the fndings were reviewed with prior techniques utilising the bilateral method's parameters of d = 1.8, r = 2 * n, and image size 7 × 7. Te adaptive bilateral flter technique was compared to earlier techniques. In this proposal, the parameter values should be set to d = 1.8, r = 1.0 * n, and the image size should be 11 × 11 at each level. Te proposed adaptive bilateral approach was found to produce better results when the MSE rate and PSNR values were compared with those of earlier methods.
To modify the various analyses of the original picture resolution in order to remove the noise from the original image. In addition, it is possible to separate picture information from noise information, which is preferred from one resolution level to another. For the various image resolutions, wavelet thresholding and the adaptive bilateral technique were combined. To minimise low-frequency noise components, the adaptive bilateral technique was used to sub-bands. Refer to Table 1, which is shown in the input image with noise and discussed with the preceding section, to choose the parameter values for the standard bilateral flter, d � 1.8 and r � 2. Table 1 demonstrates a comparison of several fltering techniques in order to identify the improved accuracy of the suggested approach. Te adaptive bilateral flter is shown in Figure 5 to have a better MSE (Mean Square Error) rate when preprocessing approaches.
Te new method's superior accuracy rate in comparison to the median flter and bilateral flter was demonstrated by a graphical view examination of earlier techniques.
A comparison of several fltering techniques is shown in Table 2 in order to identify the outcomes of the suggested method that have higher accuracy. Te PSNR rate of the adaptive bilateral flter was demonstrated by comparing the flter method.
Te new method's superior accuracy rate in comparison to the median flter and bilateral flter was demonstrated by a graphical view examination of earlier techniques shown in Figure 4.

Experiment Result for Hybrid GCPSO-FCM.
Performance of the proposed image segmentation method of hybrid GCPSO-FCM where the more accuracy rate to fnd the tumour identifcation.
Te image segmentation accuracy rate was measured using # TPR, # TNR, # FPR, and # FNR by comparing various algorithms with manual results.
In Table 3 demonstrates a comparison of several fltering techniques in order to identify the improved accuracy of the suggested approach. Te image segmentation metrics are divided into region-based and boundary-based methods. A diferent performance metrics are used to evaluate in the segmented image as TP, TN, FP, and FN.
In Table 4, we found the performance measures of the proposed image segmentation method of hybrid GCPSO-FCM where the more accuracy rate to fnd the tumour detection in Statistical analysis of FPSO-FCM Algorithm. Table 5 demonstrates a comparison of several fltering techniques in order to identify the improved accuracy of the suggested approach. Te image segmentation metrics are divided into region-based and boundary-based methods. A diferent performance metrics are used to evaluate in the segmented image as TP, TN, FP, and FN. We propose, a new Hybrid GCPSO-FCM algorithm use to each particle to every number of generations/iterations so the ftness value of all the particles to improve the Accuracy rate. Table 6 shows that the performance comparison of various segmentation algorithm PSO and FPSO-FCM accuracy rate to be compared with the proposed technique GCPSO-FCM was better accuracy rate to detect the early detection. Figure 6 shows the graphical view analysis of various segmentation algorithm PSO and FPSO-FCM accuracy rate to be compared with the proposed technique GCPSO-FCM was better accuracy rate to detect the early detection.
In Feature Extraction, using the four GLCM features that are measured above are computed using applied techniques and are examined. Te structure exhibits substantial image intensity changes using contrast measurement. Te four segmentation strategies that have been suggested ofer a maximum accuracy rate of 95% compared to the limitations of the current algorithm. Tis study uses GLCM features and a matrix to analyze statistical texture in images of brain tumours. In order to determine four directions such as 0°, 45°, 90°, and 135°, the Feature Analysis measures four texture parameters, including energy, contrast, correlation, and homogeneity with two picture intensities, 128 and 256. One of the four GLCM features, strip str2, has a high intensity of 569 for angle 45°in the average contrast, whereas angle 0°has the lowest value of the contrast.
Te correlation measures the correlation between neighbouring pixels, which spans from −1 to 1. In addition to the value of energy, the count of homogeneity aids the correlation because it is high in the 0°angle. When the image intensity is reduced, the correlation and homogeneity values do not decrease as much as the contrast value.
In addition, the energy has a high value when measured vertically. In comparison to other angles, the angle of 0°has a higher energy value. When the contrast value was decreased from 256 to 128, so was the image intensity. Even though the Sample Image   IMG_1  IMG_2  IMG_3  IMG_4  IMG_5  IMG_6  IMG_7  IMG_8  IMG_9  IMG_10 Median [5] Bilateral   IMG_1  IMG_2  IMG_3  IMG_4  IMG_5  IMG_6  IMG_7  IMG_8  IMG_9  IMG_10   0 Table 7.
In Classifcation, the fve types of kernel functions using SVM classifcation. Te polynomial kernel function is tested from 1 to 5°, and the scaling factor is tested from 1 to 5 in the kernel function of Gaussian Radial Based Function, ANOVA kernel, and Exponential Radial Based Function. Te performance of the present research method was evaluated in form of sensitivity, specifcity, and accuracy.
Te performance measures the SVM classifcation with GRBF kernel function is giving a better accuracy rate in this present research work and it also gives 98.83% of sensitivity, 90% of specifcity, and a value of accuracy of 95.32% shown in Table 8. Table 9 shows the performance comparison of various hybrid techniques to detect the early detection of brain tumours, the proposed technique GCPSO-FCM was compared and proved with a better accuracy rate to detect the early detection.
In that Table 10 shows the result analysis of various hybrid techniques to compare proved with the proposed techniques. Te proposed techniques was shows the improved accuracy rate.  IMG_1  IMG_2  IMG_3  IMG_4  IMG_5  IMG_6  IMG_7  IMG_8  IMG_9  IMG_10

Conclusion
With the fndings of this study, we ofer an efective technique for improving image quality, eliminating noise in the actual image, and locating, and limiting boundary and texture discontinuities to adjust or take control of the spatial and intensity parameters in a bilateral method that results in a higher MSE and PSNR value. Low-frequency sub-bands can beneft from the wavelet threshold and adaptive bilateral flter combination. We suggested a hybrid segmentation technique that combines GCPSO-FCM (Fuzzy C-Mean) and achieves 95.32% accuracy. Te four segmentation   Computational Intelligence and Neuroscience strategies that have been suggested ofer a maximum accuracy rate of 95% compared to the limitations of the current algorithm. Tis study uses GLCM features and a matrix to analyze statistical texture in images of brain tumours. In order to determine four directions such as 0°, 45°, 90°, and 135°, the Feature Analysis measures four texture parameters, including energy, contrast, correlation, and homogeneity with two picture intensities, 128 and 256. One of the four GLCM features has a high intensity of 569 for angle 45°in the average contrast, whereas angle 0°has the lowest value of the contrast. Te correlation measures the correlation between neighbouring pixels, which span from −1 to 1. In addition to the value of energy, the count of homogeneity aids the correlation because it is high in the 0°angle. When the image intensity is reduced, the correlation and homogeneity values do not decrease as much as the contrast value. Finally, it has been demonstrated that the GCPSO-FCM utilising an SVM classifcation methodology is suitable for categorising the image as malignant and noncancerous through the comparison of graphical view analysis of prior methods. In addition to having a 95.92% accuracy rate, SVM classifcation with the GRBF kernel function also has 99.53% sensitivity and 90% specifcity.

Data Availability
Te data used to support the fndings of this study are included within the article.