INTERNATIONAL JOURNAL OF COMPUTERS COMMUNICATIONS & CONTROL Online ISSN 1841-9844, ISSN-L 1841-9836, Volume: 18, Issue: 4, Month: August, Year: 2023 Article Number: 5186, https://doi.org/10.15837/ijccc.2023.4.5186 CCC Publications Deep learning based semantic segmentation approach for automatic detection of brain tumor S. Markkandeyan, Shivani Gupta, G. Venkat Narayanan, M Jithender Reddy, Mahmoud Ahmad Al-Khasawneh, Mohammad Ishrat, Ajmeera Kiran S. Markkandeyan School of Computing, SASTRA Deemed University Tirumalaisamudram, Thanjavur, India. drsmkupt@gmail.com Shivani Gupta School of Computer Science and Engineering, Vellore Institute of Technology Chennai, India gshivani554@gmail.com G. Venkat Narayanan St. Joseph’s College of Engineering Kanchipuram, Tamil Nadu, 600119, India gvenkatnarayanan@gmail.com M Jithender Reddy Vasavi College of Engineering (A), Hyderabad, India m.jitenderreddy@staff.vce.ac.in Mahmoud Ahmad Al-Khasawneh Department of Information Technology, School of Computing, Skyline University College, University City Sharjah Sharjah, UAE. mahmoud@outlook Mohammad Ishrat Computer Science and Engineering, KoneruLakshmaiah Education Foundation Vaddeswaram, AP, India. ishratgzp@gmail.com Ajmeera Kiran Deptartment of Computer Science and Engineering MLR Institute of Technology, Dundigal, Hyderabad, Telangana, 500043, India. kiranphd.jntuh@gmail.com https://doi.org/10.15837/ijccc.2023.4.5186 2 Abstract Initially, fromBRATS 2013 dataset the input image is acquired and is preprocessed, segmented using Convolutional neural network (CNN) based semantic segmentation, and were classified using Improved multipath GoogLeNe tCNNclassifier approach.The stage of preprocessing is carried using Bias corrected filtering. Ascheme of deep learning-based semantic segmentation of brain tumors for MRI images is proposed to classify the brain tumor effectively. In this approach, semantic seg- mentation is employed for segmentation purposes. Improved multipath GoogLeNetCNNis applied for classifying Brain tumor for classifying brain MR images and for grading the brain tumors into three classes (Meningioma, Pituitary Tumor, and Glioma). There are a total of 3064 T1-weighted contrast- enhanced pictures representing 233 patients in the input dataset.Accuracy, sensitivity, specificity, & precision estimates for the proposed method are calculated in MATLAB. The obtained results expose that the projected method achieves overall classifier performance rate of 99.7accuracy, sensitivity 100%, specificity 99.717%, and precision of 99.06%.. Results show that the suggested system outperforms state-of-the-art methods. Keywords: Brain Tumor, deep learning, artificial intelligence, semantic segmentation, MRI images, GoogLeNet CNN classification, convolutional neural network, bias- corrected filtering 1 Introduction It has long been known that cancer of the brain is particularly deadly. These tumours form in a broad array of brain areas critical to normal bodily processes. Metastases from primary tumours to other parts of the Brain evade detection by conventional imaging techniques. It may be challenging for people with brain cancer to find effective therapy because of the difficulty in identifying the origin of their disease. Nowadays, the use of artificial intelligence approaches in image processing is a significant factor that improves accuracy. Cancer rates have been climbing steadily over the last several decades. Early cancer diagnosis is notoriously challenging. Not knowing where to begin with therapy (radiation, chemotherapy, etc.) makes late cancer iden- tification extremely dangerous, if not fatal. In recent days, (Golatkar A, 2018) (N. Ravikumar, 2018)As CAD becomes more commonplace, fewer workers can do more. Digital pathology is quickly becoming a powerful diagnostic tool. Several problems in the area of processing medical images have been alleviated by the use of deep learning techniques in recent years (Amin J., 2020). A technique of classification for the classification of brain tumors was presented, depending on the deep CNN. (Carneiro G, 2017, GurusamyR, 2017., Usman K, 2017., Sajjad M, 2019., Kharrat A, 2019., Rathi V. G. P, 2015., Rathi V. G. P, 2012). CNN is often regarded as the best method for resolving classification issues when utilizing data with high dimensions like images as input. These associations "learn" to recognise and make sense of images’ local context in order to extract and categorise data. (Kharrat A, 2019., Rathi V. G. P, Rathi V. G. P, 2012). In this paper, abrain image analysis method was presented for the classificationof brain tumor type. In this technique, segmentation processing and preprocessing of images are used. Moreover, brain cancer histopathology pictures are classified using deep convolutional neural networks. • To offer a preprocessing method that improves input picture quality and removes any noise from the dataset image. • Semantic segmentation of the preprocessed picture is used in order to. • To present a Deep learning-based GoogLeNet classifier to separate the abnormal and normal stages. • To demonstrate the efficacy of the suggested scheme by estimating the proposed system’s performance. In Section II, we’ll talk about the many segmentation techniques that may be used to spot tumours in the brain. Several steps are outlined in Section III for locating a brain tumour. In Part IV, we assess the performance of the proposed technique and compare it to existing methods. In section V, an overview of the implemented approachis being presented. 2 Related Works This section provides an explicit explanation of various existing methodologies employed in breast cancer histology image analysis developed a model for the extraction of brain tumors from the 2D MRI(Magnetic Resonance brain Images) using thealgorithmFuzzy C-Means clustering, followed by both outdated classifiers and deep learning methods suggested a new CNN model termedBrainMRNet. This structural design is built on module’s attention and hyper column system; it has anenduringstructure. Initially, the image was preprocessed in the BrainMRNet. After that, this step is being converted into consideration modules with the utilization of the image intensificationapproach on behalf of the entire image. The Attention modules choose significant fields https://doi.org/10.15837/ijccc.2023.4.5186 3 of the image, and then the image was being transformed into the convolutional layers. One such considerable approach that the model of BrainMRNet employed in the convolutional layers was the hypercolumn. Using this method, the features retrieved at each stage within the BrainMRNet models are kept consistent through the final layer’s array structure. The intention was selecting the most effective and best features, along with the supported features in the array. Invasive ductal carcinoma (IDC) of cancer of the breast (BCa) in entire slide images (WSI) may be identified using a deep learning approach proposed by the BrainMRNet system, which makes use of publicly available MRI data. One method of deep learning is to simulate the learning process in a computer. This approach was motivated by the brain’s capability to process information at various levels of knowledge, or "levels," in order to predict the cancer transmissions score using breast cancer histology full slide pictures. Using molecular information or mitotic counting, this method then delivers a completely automated conclusion for cancer prediction. There are three distinct parts to the construction, and they all collaborate to provide optimal functionality. Using a combination of an image analyzing part capable of processing entire images of slide, a deep judgement-dependent detection of mitosis, as well as a propagation achieve forecasting component, we demonstrated the viability of a novel (CNN) depending prearranged regression framework for handling proceeding cells, inhomogeneous surrounding noises, and a massive variation in size and shape. The suggested approach supplied an innovative deep distributed convolutional neural network (CNN) that consists of two processes and needs just a few number of training photographs using weak observation (immediately a tick in the middle of the object). To begin, a basic retrieval strategy based on convolutional neural networks (CNN) was presented to find and localise mitotic possibilities while keeping the sensitivity high. A cross-domain transmission of information is used to refine a model of exact discrimination, which is then used to find and eliminate the brain cancer from the individual’s MRI scan images. In this broadside, a method for brain tumor separation was beingestablished on the 2D-MRI actualitiesthatconsent the tumor tissue documentation by means ofexcessivereproducibility and accuracy-related to manual processes. The brain tumor discovery’s key step is to spot the asymmetric and symmetric brain form that will label the indiscretion. On following this step, the subsequentphase is segmentation, whichwasmade on two methods Morphological operation andF- Transform (Fuzzy Transform) developed a new automatic method based on the multi-segmentation of the brain tumor region. An enhanced algorithmregion-growing was based on quasi-Monte Carlo and We developed a new method for identifying and extracting the cancer area from human head examination MRI images by combining the colour channels obtained by several modalities into a single RGB picture and utilising an expectation- maximization strategy to describe the most-preferred classes. Combining NN with a C5.0 enhanced decision tree model with multimedia color channels and key features of an MRI brain picture, the Cleveland Heart Disease database may be used to provide accurate CAD predictions suggested acompetent method for the investigation of regular and irregular states of the neuromuscular system presented a different technique of feature extraction and selection. This approach then classifies the tumour as grey matter, CSF, white issue, normal, and abnormal based on the texture, intensity, and shape-dependent aspects presented a technique for segmenting human MRI brain image at which the detection and characterization of the tumorwere considered using Wavelet packets feature spacesand HSOM. 3 Proposed Work In this part, the suggested mechanism is shown in great detail. Figure 1 depicts the overall process flow in the proposed system: https://doi.org/10.15837/ijccc.2023.4.5186 4 Figure 1: Overall flow of the proposed mechanism A. Image acquisition Initially, the images are taken with the utilization of an MRI scan, and theimagesscanned were exhibited in the 2D matrices with the element’s pixels. Then, these conditions were based on the matrix size and arena of opinion. The images are then kept in MATLAB and are exhibited as the grayscale image of about 256 * 256. The grayscale image values range from 0 to 255, with 0 viewing the 255 and total black color viewing the untainted white color. The Input among these ranges varies in the intensity that varies from white to black. The Patients betweentheages of 20 to 60 years, which covers 30 men and 30 women, wereexamined on an experimental basis. Their MRI scan images were stored in the image database in JPEG image formats. The input image is shown in figure 2. Figure 2: Input image B. Pre-processing Preprocessing occurs after a picture has been taken. When preprocessing an image, it’s common practise to remove background noise and brighten any fuzzy areas that were imported from an input database. For this filtering process, we use a bias-corrected filtering technique. In this step, we apply filters to the grayscale image and transform it to access the enhanced portion of the picture. A chime impact is utilized as well with the apparent goal of producing a superior and distinguishable end-result. Figure 3 shows both the original input and the corrected input bias pictures. https://doi.org/10.15837/ijccc.2023.4.5186 5 Figure 3: Input image, input-bias corrected Despite the current use of a filter with an, this method might be used to the processing of the kind of picture data necessary for biological research. To improve the picture, the Sobel operator is used to eliminate the high-pass filter and replace it with a mask. In MRI images, a bias field appears as a smoothed low-frequency signal due to inconsistencies in the field of magnets of the MRI equipment. High-frequency information, such as edges, lines, and intensity values, are evenly diluted when pictures are blurred owing to a bias field. Changes at this level often don’t matter in the clinic. The spatial invariance of the modified image, which is essential for precise segmentation and classification, is therefore diminished. As a result, before to engaging in segmentation or classification, a preprocessing step is required to mitigate the effect of the bias field. In the bias corrected filtering, the enhanced image is filtered by initializing the parameters by setting lambda, gamma, and theta values. The bandwidth is set for the estimation of angle. Thus, the computation of angle yx and yy is made using the formula provided in pseudocode which offers the filtered coefficient value for providing the filtered image. C. CNN based semantic segmentation After preprocessing, the image is segmented from the input pixel labeled dataset. This work aims to attain a fully automated segmentation technique depending on CNN for the extent of brain tumor segmentation from MRI images. Manually segmenting a brain tumour is a laborious process, and the quality of the results is heavily dependent on the expertise of the operator. Therefore, https://doi.org/10.15837/ijccc.2023.4.5186 6 an automated technique for segmenting brain tumors semantically is essential. The semantic segmentation is the process at which appropriate target label is assigned for each pixel with the use of region growing as a ground truth. The purpose of CNN dependent semantic image segmentation is to segment each label pixel of the image by a subset of the depicted thing that comes after the original. Due to the need of making predictions for every pixel in a picture, this process has been referred as dense prediction. The semantic segmentation is employed for recognizing the lesion region pixels from the healthy tissues that pays high attention in brain tumor regions thereby providing exact subdividing. In order to collect more instructive local characteristics, CNN is employed as the basic unit in the segmentation approach. In the CNN based semantic segmentation, the filtered image is taken as input, the parameters are initialized by setting the cut threshold range, and tumor diameter. The mask region is computed so as to estimate the tumor part and the particular segment. On estimating the size of image and the segment of image the threshold value is related to segment and recognize the affected portion of brain. The training levels and testing levels were related with the threshold limit to identify the specific portion of image. Thus, from this, the segmented outcome is found. The traditional methods of semantic segmentation all relies on the deep CNN that was superior by a huge gap to early techniques. Figure 6: Segmented outcome Figure 4is the depiction of the segmented outcome. In this, the network is trained and is evaluated, thereby inspecting the outcome of segmentation by importing the pixel labeled dataset. D. Correlation based feature extraction https://doi.org/10.15837/ijccc.2023.4.5186 7 The segmented region is taken as input for the process of features extraction approach. The feature extraction is done with the use of correlation-based feature extraction. The matrix level of the segmented portion is computed for estimating the feature level set. A Gray Level Co-occurrence Matrix (GLCM) method may be used to derive statistical texture attributes of the second order. Clusters of three or more adjacent pixels are common in third- and higher-order textures. The mathematical function GLCM may often be used to successfully eliminate the artefacts. The integrity of the original picture may also be maintained. The image might be removed for further study. GLCM may be used to estimate the pixel frequency to a given degree of precision. It is necessary to question the reliability of a single pixel & locate a replacement pixel using its path length (l) and the difference in value (m) between its neighbours. m typically takes on a single value, although this might work to the benefit in certain contexts. The obtained directional value may be used to completely remove all identifying features from the segmentation images. How closely the GLCM components are spread out along the diagonal is measured by homogeneity. Contrast may be used to assess regional shifts, whereas correlation can quantify joint probability. Energy is equal to the product of a system’s parts squared. An explanation of the GLCM method is provided below: Step 1:Pixels must be counted in pairs, with the first pixel in each pair being assigned the value I and the second pixel in each pair being assigned the value j. Step 2: Pix [i, j] is the notation for a matrix in row and column order. Step 3: In the absence of symmetries in Pix [i, j], the total amount of pixel pairs with grey levels [i,] need not be equal to the total amount of pixel pairings with grey levels [j, i]. Step 4: The normalised procedure is carried out using the pix [i, j] elements as the starting point. Simply dividing the number of pixels in each entry by the entire amount of pixels yields the normalisation. Step 5: You may get the normalised GLCM N [i, j] by, N [i, j] = (P ix[i,j])∑ i ∑ j P ix[i,j] (1) You may adjust the GLCM procedure as follows: P (m, s) = G(m, s, o, ∅)/ ∑H m=1 ∑H s=1 G(m, s, o, ∅) (2) (m,s) was the element of the m and l pixels, P represents the characteristics of a picture, and G is a frequency vector; (m,s) is a frequency of the specific component, which would typically contain the values of the pixels of l and m, ∅ stands for the constant relative to the mean. For the purpose of classification, the following characteristics were calculated and trained. The GLCM allows for the collection of many distinct characteristics. Entropy This gives broad information on image-compacting artefacts. Entropy = − ∑H−1 m=1 ∑H−1 n@s=1 nP (s, o) ∗ log(P (s, o)) (3 ) where R(s,o) represented the feature frequency The constants P and H are denoted here. Angular moment The GLCM values are determined by averaging the regions of high and poor homogeneity in the picture. There would be a lot of angular momentum but just a little precision. Typically, the photos are compared to ensure they are consistent. Angularmoment = ∑H−1 m=1 ∑H−1 n=1 P (n, o) 2 (4) Contrast Quantitative measures of visual quality are determined. Disparities between regions are often examined. Contrast = ∑H−1 s=0 s 2(5) Inverse Difference Moment It’s also a generic tool for gauging levels of similarity. IDM = ∑H−1 m=1 ∑H−1 s=1 1/1 + (s − o) 2P (s, 0)(6) Energy It might be used to figure out how many squares of the return are realistic. E = ∑H−1 0=1 ∑H−1 s=1 P (s, o) 3 (7) https://doi.org/10.15837/ijccc.2023.4.5186 8 Variance It is possible to directly compute the standard deviation of the grayscale values. V AR = ∑H−1 m=1 ∑H−1 s=1 (P (s, o)) 2 − ∅2 (8) Sum average In most cases, we can calculate the frequency of the links between pixels.∑2H−1 s=0 nPx+y(l) (9) The outcome of these features will help in the identification of cancer presence along with their types like ma- lignant and benign. The value of the contrast, area, correlation, and entropy was higher for benign cases on comparing malignant ones. However, the amount of homogeneity is supposed to be lower for benign than the malignant tumor. E. Improved Multipath GoogLeNet CNN classification After the segmentationprocess, there is a necessitate for differentiating cancer cells, whether it is the malig- nant or benign stage. For the purpose of classification, GoogLeNet CNN was employed, one such algorithm that was well-established. The GoogleNet CNN classification architecture comprises of 6.8 million parameters nine inception modules, To do so, we’ll use a network with the following layers: five highest pooling layers, a dimensionality-reduction convolution layer, two twisted layers, a single fully-linked layer, two normalization layers, one middle pooling layer, and finally a linear layer triggered by the judgement output. Four convolutional layers are dedicated to dimensionality reduction, and there are six layers for maximum pooling as well as one convolutional layer of different types in each inception module. The activation function of ReLU is applied uniformly, while dropout regularization is employed in the completely connected layers. It also offers improved accuracy over AlexNet. The target is being given a score based on the likelihood scale. It is the algorithm for convolution which is pre-trained.GoogLeNet CNN permits the variations here between single variable dependent and more such independent factors to be evaluated. The CNN makes calculations based on its best guesses. There is just one overall distribution. In this technique, CNN was used to read and reformat the picture before doing the classification based on a probability estimate. CNN, developed by Google, is a deep neural learning network. When it comes to visual recognition and classification, GoogLeNet CNN represents a huge step for- ward. In characterizing and categorizing pictures, they are more often utilised to break conceptual continuity and to run optics. The GoogLeNet has been used to create layers. •CNN. •ReLU •Convolution •Pooling •Completely connected When compared to other image classification techniques, GoogLeNet CNNs need much less preprocessing procedures. This CNN will serve specialised needs across a variety of industries. Convolution Highlightning of the image data is the key function of this convolution phase. The convolutionary layer in Google NET CNN is consistently the first step. The features will be identified, and the characteristic map generated from the input picture ReLU layer The next step is the ReLu layer. In this case, the enactment method has been extended to the feature maps to improve network non-linearity. In this scenario, negatives will easily be omitted. Pooling The Pooling method will reduce the input size slowly. This will speedily evaluate the necessary parameters by growing the amount of parameters needed. Flattening It is a straightforward move, in the sequential column of numbers, where the polled function chart should be flattened. Completely connected layer Here we may merge the functionality with the attributes. The classification method can be done with the high percentile of precision. The error will mainly be measured and recorded. https://doi.org/10.15837/ijccc.2023.4.5186 9 Figure 5: GoogleNet CNN Architecture Softmax: To convert the unstandardized network performance across predicted output groups into a probability distribu- tion, neural networks often use the softmax method. The Softmax can be used to solve a wide range of problems in many academic disciplines. According to these odds, 1.0 would be shown. Keep in mind the associated Soft- max variants: Softmax is Complete and will measure a likelihood for increasing possible type. Softmax measures a chance, but only for a specific case of negative names with all positive names. This CNN allows for the comparison of two or more values to determine how much of a difference there is. CNN uses a probability calculation and a function. This is the total amount doled out. CNN may first read & redistribute the picture, and then evaluate the classification process based on its class probability. By stimulating a neuron, we kick off the process, alj = σ ∑ K Wj k lalK + b l j (10) A vectorized version of the equation is possible, al = σ(wlal−1 + bl) (11) In which quadratic sets may the training set be combined, C = 1/2ĘY − alĘ2 = 1/2 ∑ j (yj − a l j ) 2 (12) The gradient output is given by, (∂c)/(∂wkjl) = akl−1δlj=C (13) feature (F) and pointed feature (P) constitute the characteristics that have been categorised. They must be reported as The features of data, which is chosen for training the classifier, that is from the Use of a GLCM (Gray Level Co-occurrence Matrix) to analyse data. The data gleaned from the tested photos for categorization purposes during training. Here the classification is performed by GoogLeNetclassification.The feature extracted outcome is trained to classify the image. The batch size value is set to 10 by means of the primary learn rate of 1e 4 and the maximum epochs of 20. The best epochs number is being altered as per the validation criteria along https://doi.org/10.15837/ijccc.2023.4.5186 10 with validation frequency of 300 iterations.The network layers and train features were initialized by initializing train label as 70%, and test label as 30%. Thus, the training and testing ratio are 70:30. For the trained and tested features, the prediction is made by their predicted label estimation. Glioma, meningioma, and pituitary tumours are the three broad categories into which brain tumours fall. The confusion matrix for the proposed methodology is shown below: Table 1 confusion matrix of propose method Meningioma Glioma Pituitary tumor Meningioma 1031 0 1 Glioma 0 1069 0 Pituitary tumor 0 3 960 Table 1. Shows the confusion matrices which in turn precises scheme’s performance. The signifies the predicted values (output of system) and the true labels (ground truth). The algorithm for GoogLeNet CNN is shown below: In this method, the classification operation was made by the estimation of class probability, which in turn displays the labeled image. Thus, classification of types of tumor grades namely Meningioma, Glioma, or pituitary tumor were made. 4 Performance Analysis Results and analysis of the suggested study’s performance analysis are shown below: the proposed study is automaticbrain tumor detection in brain images. A. Dataset Description The data used for the training model was collected between 2005 and 2010 by researchers at Nanfang Hospital & General Hospital of Tianjing Health college, China, and made publicly available online in many different forms between 2015 and 2017. There are 233 cases included in the database, and they were all acquired using contrast-enhanced T1-weighted imaging of meningioma, glioma, or hypothalamic tumours. Brain tumours may vary greatly in terms of their size, kind, and location[24]. Table 2 summarises the number of patients and the number of slices for each kind of brain tumour in the dataset. https://doi.org/10.15837/ijccc.2023.4.5186 11 Table 2 Dataset showing number of patients and slices Type of tumor Number of pa- tients Number of slices Glioma 91 1426 Meningioma 82 708 Pituitary tumor 60 930 233 3064 B.Performance analysis of the proposed mechanism Performance and accuracy were examined, and the outcomes were shown below. The comparison of the proposed outcome is related to the existing technique to prove the proposed mechanism’s effectiveness. Accuracy It is a statistical measure of a person’s inclination towards mathematics, with accuracy defined as the fraction of observations that are really correct (both negative and positive). Accuracy (A) = (TP+TN)/(TP + TN + FP + FN) (18) Sensitivity Sensitivity, also known as the true positive rate, the likelihood of detection, and recall, is a measure of how many true positives can be reliably detected. Sensitivity=TP/ (TP+FN) (19) Specificity The term "specificity" refers to the "true positive rate," which provides a rough approximation of the genuine negative quantities that were accurately recognised. Specificity = TP/(TP+FP) (20) Precision The number of correctly labelled photos is an example of the output parameter known as "precision." A helpful reminder will be the total number of instances that fall inside the various categories. The ratio of two images provides an accurate measurement of precision. Precision = (TP)/(TP+FP) (21) Segmentation analysis The current method [25] presents multi-cascaded convolutional neural networks (MCCNN)-based brain tumour segmentation, which boasts a sensitivity rate of 0.9074 and a specificity rate of 0.9918. The computation time of the MCCNN is 1.5-3 mins. Whereas, in presented approach, the computation time (29.133 sec) which is lower than existing technique. Below, we provide a comparison study of the segmentation system we suggest: Table 3 Comparative analysis of the proposed segmentation system Factors CurrentScheme SuggestedScheme Sympathy 0.9074 0.9153 Specificity 0.9918 0.9928 Calculation time 1.5-3 mins 29.133 sec UntrueConfident Rate 2.24 0.0072 UntrueBad Rate 1.84 0.0847 Comparison of the current segmentation result with the suggested one is shown in Table 3. When compared to other segmentation algorithms, the results obtained demonstrate that the suggested scheme is superior. D. Classifier performance The current method [24] uses a deep neural network to offer multi-classification of brain tumour pictures. With a total accuracy of 96.13%, the current system architecture produces a notable result. However, the suggested method has an approximate 99.7 percent success rate.The accuracy rates of the suggested approach, the state-of-the-art method, and the combination of the two are compared and contrasted. https://doi.org/10.15837/ijccc.2023.4.5186 12 Figure 6 Accuracy vs. No of Iterations Figure 7 Sensitivity vs. No of Iterations Figure 8. Specificity vs. No of Iterations https://doi.org/10.15837/ijccc.2023.4.5186 13 Figure 9. F-score vs. No of Iterations Figure 6, 7, 7 and 9 shows the results of an estimated comparison of the suggested and current method for detecting Glioma. The study shows that the presented scheme is superior than existing techniques. The sensitivity is 98.85%, specificity of about 98.97%, precision of 98.85%, and accuracy of 98.91%. A look at how the suggested method stacks up against the status quo was estimated and is represented in figure 7, which in turn detects the presence of the pituitary tumor. The research confirms that the proposed method outperforms the alternatives. The sensitivity is 97.26%, specificity of about 99.00%, precision of 96.82%, and accuracy of 98.58%. The accuracy, sensitivity, specificity, & precision of the proposed and current methods for diagnosing brain abnormalities such meningioma are compared and contrasted in Figure 8. The results are as follows: 97.71 percent sensitivity, 99.5 percent specificity, 98.61 percent precision, and 99.13 percent accuracy.The results of the study proved that the recommended approach is superior than competing ones. Table 4 Overall classifier outcome Parameters Existing system Proposed system Accuracy 96.13 99.78 Sensitivity 95.03 100 Specificity 96.9 99.717 Precision 96.06 99.06 Table 4is the analysis of comparison between existing and proposed classifier overall accuracy, sensitivity, specificity, and precision. Compared to state-of-the-art classifiers, the results demonstrate that the suggested method performs better. F ¯ igure 10. Validation accuracy of the classifier varied by number of iterations https://doi.org/10.15837/ijccc.2023.4.5186 14 F ¯ igure 11. Performance estimation of loss for number of iterations Classifier validation accuracy as a function of iteration count is shown in Figure 9 of the performance analysis. Likewise, figure 10 is the performance analysis of loss for each varying iteration. The overall outcome of the classifier is as follows: the proposed technique offers overall accuracy of about 99.7%, sensitivity 100%, specificity 99.717%, and precision of 99.06%. Therefore, the proposed system is better in both segmented and classifier outcome than other existing techniques. 5 Conclusion Automatic brain tumour segmentation and classification using a deep learning based AI method is described. The findings demonstrate that a tumour may be properly segmented using Semantic segmentation & GoogLeNet CNN Classification in a MATLAB setting. The purpose of CNN dependent semantic image segmentation is to segment each label pixel of the image by a subsequent class. Improved multipath GoogLeNet CNN was employed in this approach, which is an effective mechanism for the classification and detection process. This is being employed for the classification of object depending on their probabilities. This is termed as the algorithm that were pre-trained convolutional one.Furthermore, the outcomes of this projected study out comes produce clear proportions regarding the abnormal and normal brain tumor region segmentation that, in turn enables the physicians to have reliable method. The performance of proposed technique is estimated using MATLAB In regards to how precise, sensitive, and accurate it is. The obtained results expose that the projected method achieves overall classifier performance rate 99.7% accuracy, sensitivity 100%, specificity 99.717%, and precision of 99.06%. The achieved results show that the suggested approach is superior to other methods currently in use. To categories brain MR images and categorise tumours into Meningioma, Pituitary Tumor, & Glioma, an enhanced multiple paths GoogLeNet CNN is used. The obtained results expose that the projected method achieves 3.65% improvement in accuracy, 4.97% in sensitivity, 2.817% specificity improvement, and 3% preci- sion improvement than existing method. Acknowledgement There is no acknowledgement involved in this work References [1] Ottom, M.A., Rahman, H.A., & Dinov, I.D. (2022). Znet: Deep Learning Approach for 2D MRI Brain Tumor Segmentation. IEEE Journal of Translational Engineering in Health and Medicine, 10, 1-8. [2] Archana, K.V., & Komarasamy, G. (2023). A novel deep learning-based brain tumor detection using the Bagging ensemble with K-nearest neighbor. Journal of Intelligent Systems, 32. [3] Chetty, G., Yamin, M., & White, M. (2022). A low resource 3D U-Net based deep learning model for medical image analysis. International Journal of Information Technology, 14, 95 - 103. [4] Mohan, P., V E, S., Subramani, N., Subramanian, M., & Meckanzi, S. (2022). Handcrafted Deep-Feature- Based Brain Tumor Detection and Classification Using MRI Images. Electronics. https://doi.org/10.15837/ijccc.2023.4.5186 15 [5] Chetty, G., Yamin, M., & White, M. (2022). A low resource 3D U-Net based deep learning model for medical image analysis. International Journal of Information Technology, 14, 95 - 103. [6] Aleid, A., Alhussaini, K., Alanazi, R.S., Altwaimi, M., Altwijri, O., & Saad, A.S. (2023). Artificial Intelli- gence Approach for Early Detection of Brain Tumors Using MRI Images. Applied Sciences. [7] Alqazzaz, S., Sun, X., Nokes, L., Yang, H., Yang, Y., Xu, R., Zhang, Y., & Yang, X. (2022). Combined Features in Region of Interest for Brain Tumor Segmentation. Journal of Digital Imaging, 35, 938 - 946. [8] Yang, Z., & Li, S. (2022). Dual-path network for liver and tumor segmentation in CT images using Swin Transformer encoding approach. Current medical imaging. [9] Jyothi, P.,& Singh, A.K. (2022). Deep learning models and traditional automated techniques for brain tumor segmentation in MRI: a review. Artificial Intelligence Review, 56, 2923 - 2969. [10] GolatkarA., D. Anand, and A. Sethi. 2018. "Classification of breast cancer histology us- ing deep learning." InInternational Conference Image Analysis and Recognition, 2018.837-844. https://doi.org/10.3389/fgene.2019.00080 [11] Vesal S., N. Ravikumar, A. Davari, S. Ellmann, and A. Maier. 2018. "Classification of breast cancer histology images using transfer learning." InInternational Conference Image Analysis and Recognition, 812-819.https://doi.org/10.1155/2022/8904768 [12] Amin J., M. Sharif, N. Gul, M. Yasmin, and S. A. Shad. 2020. "Brain tumor classification based on DWT fusion of MRI sequences using convolutional neural network." Pattern Recognition Letters. 129: 115-122.https://doi.org/10.1016/j.patrec.2019.11.016 [13] Carneiro G., Y. Zheng, F. Xing, and L. Yang. 2017. "Review of deep learning methods in mammography, cardiovascular, and microscopy image analysis." In Deep Learning and Convolutional Neural Networks for Medical Image Computing, Springer.11-32.DOI:10.1007/978-3-319-42999-12 [14] GurusamyR., and V. Subramaniam. 2017. "A machine learning approach for MRI brain tumor classifica- tion." Computers, Materials and Continua. 53: 91-108.https://doi.org/10.3970/cmc.2017.053.091 [15] Usman K., and K. Rajpoot. 2017. "Brain tumor classification from multi-modality MRI using wavelets and machine learning." Pattern Analysis and Applications. 20: 871- 881.https://doi.org/10.1016/j.jocs.2018.12.003 [16] Sajjad M., S. Khan, K. Muhammad, W. Wu, A. Ullah, and S. W. Baik. 2019."Multi-grade brain tumor classification using deep CNN with extensive data augmentation." Journal of computational science. 30: 174-182. https://doi.org/10.1016/j.jocs.2018.12.003 [17] Kharrat A., and N. Mahmoud. 2019. "Feature selection based on hybrid optimization for magnetic resonance imaging brain tumor classification and segmentation." Applied Medical Informatics. 41: 9- 23.DOI:10.1007/978-3-319-76354-530 [18] Rathi V. G. P., and S. Palani. 2015. "Brain tumor detection and classification using deep learning classifier on MRI images." Research Journal of Applied Sciences, Engineering and Technology. 10: 177-187. DOI 10.1088/1742-6596/1916/1/012226 [19] Rathi V. G. P., and D. S. Palani. 2012. "A novel approach for feature extraction and selection on MRI images for brain tumor classification." In Int Conf Comp Sci Eng Appl. 225-234.DOI:10.5121/csit.2012.2224 [20] Shah M. F. M., 2019. "Brain Tumor Detection using Convolutional Neural Network." Ahsanullah University of Science andTechnology https://doi.org/10.15837/ijccc.2023.4.5186 16 Copyright ©2023 by the authors. Licensee Agora University, Oradea, Romania. This is an open access article distributed under the terms and conditions of the Creative Commons Attribution-NonCommercial 4.0 International License. Journal’s webpage: http://univagora.ro/jour/index.php/ijccc/ This journal is a member of, and subscribes to the principles of, the Committee on Publication Ethics (COPE). https://publicationethics.org/members/international-journal-computers-communications-and-control Cite this paper as: Markkandeyan, S.; Gupta, S. ; Venkat N., G.; Jithender Reddy, M ; Ahmad Al-Khasawneh, M.; Ishrat, M.; Kiran A. (2023). Deep learning based semantic segmentation approach for automatic detection of brain tumor, International Journal of Computers Communications & Control, 18(4), 5186, 2023. https://doi.org/10.15837/ijccc.2023.4.5186 Introduction Related Works Proposed Work Performance Analysis Conclusion