A Review on Multiclass Brain Tumor Detection using Convolutional Neural Network and Support Vector M

A Review on Multiclass Brain Tumor Detection using Convolutional Neural Network and Support Vector Machines

ABSTRACT:

The role of clinical diagnosis in modern healthcare has grown in importance. Brain cancer is a major concern in the realm of medical imaging since it is the worst illness in the world. Brain tumour evaluation and prognosis may benefit from magnetic resonance imaging-based early and accurate diagnosis. Medical images need to be identified, segmented, and categorised so that radiologists may use computer-aided diagnostic techniques to help them find brain tumours. Radiologists find the process of manually detecting brain tumours to be tedious and prone to errors; as a result, an automated technique is urgently needed. Therefore, the approach for accurately detecting and categorising brain tumours is introduced. In terms of the materials and methods used, There are five stages to the procedure suggested. Initial steps include applying a linear contrast stretching to the original picture in order to locate the image's edges. In the second stage, a deep neural network architecture tailored specifically to the task of brain tumour segmentation is created. Third, we employ transfer learning to train a modified version of the MobileNetV2 architecture for feature extraction. Lastly, the best features were chosen using a multiclass support vector machine (M-SVM) and a controlled entropy-based technique. Finally, meningioma, glioma, and pituitary pictures are categorised by using M-SVM for brain tumour classification.

KEYWORDS

Biomedical image processing; brain tumor; deep learning; linear contrast stretching; segmentation.

I.INTRODUCTION:

Primitive costs for patients with brain tumours are currently the highest of all cancer types. Brain tumours may develop in persons of any age due to the rapid growth of certain cell types. A brain tumour is an abnormal growth of tissue that develops in the brain or central spine and disrupts normal brain function [1]. These big tumour cells may be classified as malignant (cancerous) or benign (non-cancerous) based on their location, size, and surface area [2]. Primary and secondary tumour sites refer to the most recent developments of cancer cells. The initial tumour region is defined as the stage at which cancer cells are deemed benign. Tumors that originate in brain tissue are consideredprimarybraintumours,andtheyarecurable or at least manageable with the right treatment. Malignancies that originate in other organs and spread to the brain are called secondary tumours. Only via proper surgery or radiation treatment [3] can the pretentious patient's tumour be treated. Due to the danger posed by brain tumours to healthy brain tissue, monitoring their development is crucial for patient survival[4].

Tumorscalledmeningiomasmayspreadtothebrainand spinal cord. The tumours themselves are composed of three meningeal layers [5]. Meningiomas often manifest asasymmetricallobarmasseswithsharpmargins[6].

Survival rates for people with meningioma depend on factorssuchastumoursize,wherethetumourislocated, and the patient's age. Symptoms of meningioma include obsessive clinginess, frequent headaches, and limb weakness. Tumors of benign meningiomas are less than 2 mm in diameter, but those of malignant meningiomas maybeaslargeas5cmindiameter[7].Mostmalignant meningiomasarecurableifdetectedandtreatedearly.

Sinceseveral differenttypes ofMRImaybe usedforthe detection of brain cancer, magnetic resonance imaging (MRI) has become one of the most common techniques for making this diagnosis [8]. Since brain tumours are potentially life-threatening, it is crucial that they be diagnosed and treated correctly. Detecting the illness in itsearlieststageswithawholebrainscanistheonlyway toavoidputtingpatientsat risk.Differenttypesof brain tissue may be detected by various MRI techniques, each of which has its own unique settling time [9]. Due to their unpredictable form and location, brain tumours might be difficult to identify using only one MRI modality. When trying to locate tumours, contrasting data from several MRI techniques is crucial [10]. T1weightedMRIcanbeusedtodifferentiatetumourtissue from healthy tissue, T2-weighted MRI can be used to outline areas of edoema, resulting in clear image areas, T4-Gd MRI can show a bright signal at the tumour edge whencontrastenhancementisused,andFLAIRMRIcan

use water molecules to suppress signals to differentiate cerebrospinalfluid(CSF)fromareasofedoema.

Calculating area, determining uncertainty in segmentation area, and tumour segmentation are challenging activities [11] because of the structural complexity and unpredictability of brain tumours, high volatility, and intrinsic features of MRI data, i.e., fluctuation of tumour size and form. Creating manual tumour segmentation is a time-consuming operation, and doctors may notice variations in segmentation outcomesowingtovariancesintumourformandshape. Meningiomas, on the other hand, are rather easy to separate, whereas gliomas and glioblastomas are more challenging [12]. It is crucial, then, to provide an automated segmentation approach to facilitate this laboriousprocess.

Brain tumour detection and monitoring need a lot of time and human error is common when done manually [13]. We have to find a way to replace the manual processeswithanautomatedone.Itisincompatiblewith brain tumour detection processes to use the present approaches, which rely on labelling methods to identify sick regions in the brain and cannot detect internal peripheral pixels.WefavourMRIoverCTscans because ofthecontrastagent'sabilitytoclearlyshowtheaffected region (CT). Therefore, MRI modalities are utilised in a widevarietyofapproachestodiagnosebraincancer.

Manydifferentapproacheshavebeenproposedinrecent years for the automatic classification of brain tumours; these can be broadly categorised as either Machine Learning (ML) or Deep Learning (DL) approaches, depending on their focus on feature fusion, feature selection,ortheunderlyinglearningmechanism.Feature selection and feature extraction are cornerstones of classification in ML approaches [14,15]. On the other hand, deep learning techniques may learn by manually extracting characteristics from photos. The latest deep learning(DL)techniques,convolutionalneuralnetworks (CNNs)inparticular,boastimpressiveprecisionandare widelyusedforMRIanalysisandotherformsofmedical imageanalysis[16,17,18].Althoughthesedrawbackscan be mitigated through the use of transfer learning [19], they still exist when compared to traditional ML approaches and include the need for a large training dataset, high complexity of time, low accuracy for applicationswhereonlysmalldatasetsareavailable,and expensive GPUs that eventually increase the cost to the user. Furthermore, knowing enough about various parameters, training techniques, and topologies might makepickingtherightdeeplearningmodelseemlikean overwhelming undertaking. Support Vector Machine (SVM), Random Forest (RF), fuzzy C-mean (FCM), Convolutional Neural Network (CNN), Nave Bayes (NB), K-Nearest Neighbor (KNN), Sequential Minimal Optimization (SMO), and Decision Tree are just some of

the machine learning-based classifiers that have been used for brain tumour classification and detection (DT). With reduced computational and spatial complexity, the CNN implementation is easier to implement. The little datasetneededfortraining,thelowcomputingcost,and the simplicity of adoption by untrained persons have attractedalotofacademicinteresttotheseclassifiersin general.

These are some of the projected improvements brought about by the novel approach of segmenting and classifyingbraintumours.

In this pre-processing stage, we apply a linear contrast stretching technique to enhance the original image's edgefeatures;

Developed from scratch 17-layer convolutional neural network (CNN) architecture specifically designed for braintumoursegmentation;

Toget therequireddatasetsfordeep feature extraction, we used transfer learning from a modified version of MobileNetV2;

To achieve this, we use an entropy-controlled technique forselectingfeatures,wherethebestfeaturesarechosen according to the entropy value. A multi-class SVM classifierisusedtocategorisethefinal characteristics; a thorough statistical analysisand comparison with stateof-the-art approaches are carried out to validate the stabilityoftheproposedmethodology.

II. RELATED WORKS

Modern medical practises often use MR imaging for the diagnosis of brain cancer [8,14]. This part takes a close look at how well brain tumours are known to be detectedandcategorised.

Numerous studies on brain tumour detection, segmentation, and classification have been conducted in recent years. Even after many publications emphasising its significance [20,21,22], the medical community continues to stress the topic's relevance. This study presents a methodology for the identification and delineation of brain tumours. Differentiating brain pictures may be done using either generative or discriminating techniques [17,23] for the diagnosis of brain cancers. Brain tumour identification using fuzzy logic and the U-NET CNN architecture was shown by Maqsood et al. [4]. It included using contrast enhancement, a fuzzy logic-based edge detection approach, and U-NET CNN classification. For this purpose, we first apply a contrast enhancement techniquetothesourcepicturesforpre-processing,then we use a fuzzy logic-based edge detection technique to find the edges in the contrast-enhanced images, and

lastlyweapplyadualtree-complexwavelettransformat different scale levels. In order to differentiate between meningioma and non-meningioma in brain imaging, features are derived from deconstructed sub-band images and then classified using the U-NET CNN classification approach. The suggested technique outperformed several state-of-the-art algorithms by a widemargin(98.59%).

Sobhaninia et al. [24] trained a CNN model for segmentation using brain MRI using a LinkNet network, combining several views to improve the model's performance; the final result was a dice score of 0.79. Johnpeter et al. [25] use an adaptive neuro-fuzzy inference classification approach to identify and locate cancers in brain MRI, yet this network seems to be rather sophisticated. Without employing edge detection onbrainpictures,ourstrategyenhancedtumourregions using histogram equalisation. The resulting work was accurate98.80percentofthetime.

Withtheuseofthemoduloandhypercolumntechnique, Togacaretal.[26]createdanetworkcalledBrainMRNet. The original photos were first subjected to preprocessing, and then they were sent to the attention module.Theconvolutionallayerreceivesthepictureand processes it based on the attention module, which controlsthefocal regionsof theimage. TheBrainMRNet model's convolutional layers make extensive use of the hypercolumn strategy. By using this technique, we were able to increase accuracy to 96.05%, since the characteristicscollectedfromeachsuccessivelayerwere savedinthearraytreeofthefinallayer.Toclassifybrain tumours, Kibriya et al. [27] established an approach based on the fusion of many features. To solve the data issue, we first apply the minimum-maximum normalisation approach to the original photos and then use enormous data extension to the pre-processed images. The final output and accuracy of 97.7% are the result of a combination of a support vector machine (SVM) and a k-nearest neighbour (KNN) classifier trainedondata fromtheGoogLeNetandResNet18 deep CNN models. Brain tumours may be detected and classified using a CNN developed by Sajjad et al. [28]. Theauthorswereabletoachievea94.58%accuracyrate by using a Cascade CNN algorithm for segmenting the tumours in the brain and a fine-tuned version of VGG19 for classifying them. Shanthakumar [29] employed watershed segmentation on brain MRI scans to localise tumours. The accuracy of tumour segmentation was improved to 94.52% using this segmentation approach, whichmakesuseofapresetlabelling schemetoachieve this result. Separating tumour regions in MR images of thebrainisshowntobepossiblebyPrastawaetal.[30]. Althoughthisapproachhasan88.17%successrate,itis onlyabletoidentifytheoutside,abnormalboundariesof thetumourarea,andnottheinteriorboundary.

For the purpose of classifying brain tumours, Gumaei et al. [31] suggested a hybrid feature extraction technique based on a regularised extreme learning machine (RELM). Preprocessing is performed using the min-max normalisation contrast enhancement approach, feature extraction is performed using a hybrid PCA-NGIST method, and brain tumour classification is performed usingtheRELMmethod.Theoverallprecisionofthisjob was 94.23%. When used to contrast-enhanced magnetic resonance imaging (CE-MRI), a fine-tuned pre-trained VGG19 model improved outcomes for Swati et al. [32], who reported an average accuracy of 94.82%. After addressing the issue of overfitting using the ResNet50 CNNmodelandglobalaveragepooling,Kumaretal.[33] suggested a brain tumour technique with an average accuracy of 97.48%. These game-changing advances have garnered widespread attention in the field of medical image analysis. Brain picture categorization by machine learning and an understanding of brain architecturewasproposedbyVeeramuthuetal.[4].

Decomposingthepictureandsubsequentlyextractingits characteristicsismadeeasierwiththeaidofMultiLevel Discrete Wavelet Transform. In order to classify the severity of the illness in the brain picture, a PNN-RBF training and classification approach is used. The hybrid strategy was created by Sanjeev et al. [5]. This mixed method utilises discrete wavelet transformation (DWT) tofilteroutirrelevantinformation,ageneticalgorithmto narrow the feature set, and a support vector machine (SVM) to classify different types of brain tumours. To enhance the effectiveness of motor imagery categorization,Gopaletal.[6]suggestedastrategybased on feed forward backpropagation of the neural network (FFBPNN). Artificial neural networks (ANN), fuzzy clustering techniques (FCM), support vector machines (SVM), decision trees (DT), K-Nearest Neighbors (KNN), and Bayesian classification are only some of the approaches available for classifying medical pictures. This ANN, as well as SVM and KNN, are examples of supervised learning methods. Unsupervised learning methods, such K-means clustering and the SelfOrganizing Map, are another category that may be used togroupdataintomeaningfulgroups.

The shift from manually crafted features to machinelearned ones has been slow, however. Several methods for learning features were in use prior to AlexNet's breakthrough. Themethods will be thoroughlyanalysed by Bengio et al. [7]. Main component analysis, picture patchclustering,anddictionarytechniquesarejustafew examples. At the conclusion of their study, in a section titledGlobalTrainingofDeepModels,Moosaetal.[8]will deploy CNNs that are taught from beginning to end. In this review, we focus on these foundational models and leave out the more surface-level ones. using traditional feature learning techniques on medical photos. Refer to

Ravi et al. (2017) [9] for a more comprehensive review of deep learning's role in health informatics, including a shortdiscussionofitsapplicationtotheinterpretationof medicalimages.

Shen et al. (2017) [10] revealed the results of a deep learning-specificstudyofmedicalpictureanalysis.While they certainly cover a lot of ground, we believe that certain crucial points have been overlooked. Medical image segmentation is essential for prompt treatment planning, and it is used for the categorization and identification of brain tumours from MR images. MethodsforMRIClassification

There are a lot of brain tumours. Imaging of the brain withMRI(MagneticResonanceImaging)isthestandard diagnosticprocedureforevaluatingtumoursinthehead. Traditional machine learning methods often assign a classification to a brain tumour based on an arbitrary attribute or the discretion of a radiologist. In this study, we use ensemble modelling using the SVM & CNN classifier [5] on MRI scans of the brain to distinguish betweenbenignandmalignanttumours.

Inaddition,whenitcomestorecognisingbraintumours, threshold-based segmentation management causes blurrededgesandlimits.

Using Resnet-50 and TL, a deep learning model was createdfordetectinganddiagnosingbraintumours.The accuracy rate of their experiments is 95%. Researchers used block-wise based transfer learning to accomplish fivefold crossvalidation. 95 percent accuracy Their technology (CEMRI) was put to the test using a benchmark dataset constructed from T1-weighted MR images.ClassifyingMRIscansofthebrainusingGoogle's neural network architecture. Classification accuracy of 98% was achieved. As a classifier, a support vector machine-based method is utilized [7]. Classification and featureextractionaretwoapplicationsofCNN.Thereare twoconvolutionallayersandtwofullylinkedlayersused in this structure. In a study using nine deep learning models, they employed Transfer Learning to identify braintumourswithaTLaccuracyof97.39%[7].

In order to delve into MR data, they switched to deep learningmode.ClassificationofMRIscanswassuccessful to the tune of 98.71 percent using the suggested approach. Despite the small size of the research, the findingswerestriking.CNN'sblueprintswerespot-onin every respect. VGG also achieved 96 percent accuracy, while ResNet50 achieved 89 percent and InceptionV3 achieved89percent.Accuracyof75%[8]Asreportedby CNN,modernbuildingsaredesignedtoworkatlightning speedswhile maintaininga 98.24percentaccuracyrate. Multi-scaleanalysisofMRIscansofbraintumoursusing CNN is highly recommended. They tested the suggested modelontheMRIimagedataset,andfoundthatithada

97.3 percent accuracy in classifying the images [9]. In ordertoclassifybraintumours,theCNNmodelusestwo convolutional layers and two fully connected layers to gather relevant data for feature extraction. Classifying braintumours,theyhada97%successrate[10].

Researchers employed a convolutional neural network ResNet34 model and a transfer learning approach to categorise MRI scans of the brain into healthy and diseased categories [20]. They used a method for improving data photographs of brain tumours are normalornottoincreasethenumberofshotsandgetto 100%accuracy[11].

The ANN model was combined with the optimization strategy oftheGrayWolfOptimizer(GWO).UsingGWOANN,theywereabletoachievea98.91percentaccuracy intheirclassifications.TheyshowedoffaResNet-50-and brainMR-traineddeepCNNnetwork [12].Withthehelp of the recommended data improvement technique, the model's accuracy increased to 97.48%. Using an MRI of the brain as training data, a Capsnet CNN model was proposedwith90.89percentaccuracy[13].Anensemble model consisting of three separate convolutional neural networkclassifiersreached98%accuracy[14].

The research found that transfer learning might be utilised to categorise brain tumours. A CNN with one of four possible architectures DenseNet-2, VGG-16, VGG19, or ResNet-50 was chosen for this task. In this investigation,weusedFigureShare.using3064magnetic resonance imaging (MRI) data to differentiate between three different types of brain tumours The produced modelwasrefinedwiththehelpofapublictestbed.The resultsdemonstratedthatthepubliclyavailableFigshare datasetpromotedknowledgeexchange.

ThecreationoftheResNet-50modelwasatriumph. Usuallywithin99.02percent.Theteampickedonwhere they left off in 2020, continuing to use the same dataset in an attempt to improve the diagnostic accuracy for brain cancer. CNNs are suggested to include two convolution layers for feature extraction and two fully linked layers for classification [16]. A total of 97.39% of casesofbraincancerwerecorrectlydiagnosedusingthis CNN system. Using the data at hand, researchers were able to identify several subtypes of brain tumours. A variety of classifiers including KNN, ANN, RF, and LDA wereapplied.Accuracyof95.56percentwasattainedby combining the KNN model with the NLBP feature extraction method [17]. When working with a brain to identify and classify tumours, the shortcomings of the aforementioned methods of transfer learning intrusiveness,complexity,andsusceptibilitytosampling errors must be overcome. Systematically, there is a dearth of research on the effectiveness and dependability of such methods. This has led to the

development of transfer learning models for use in the diagnosis and classification of malignant brain tumours. Imageswereclassifiedusingdeeplearning,auniqueand powerfulclassificationapproach,andtumourtypeswere classified using faster region-based CNN (faster R-CNN). Khairandish et. al. [1] provided an explanation of how brain tumors actually behave, and with the aid of many methodologiesandtheanalysisofresearchstudiesusing a variety of criteria, it offers a clear image of this stage. The examination is conducted in relation to the dataset, proposedmodel,proposedmodelperformance,andtype ofmethodusedineachpaper.Between79and97.7%of the publications under study had accurate results. They employed Convolutional Neural Network, K-Nearest Neighbour, K-Means, and Random Forest algorithms, in thatsequence(highestfrequencyofusetolowest).Here ConvolutionalNeuralNetworkgavethehighestaccuracy of around 79-97.7% Someswararao et. al. [2] developed a new novel method for detecting tumors in MR images By using machine learning techniques, particularly the CNN model, in this study. This study combined a CNN model classification challenge for determining whether or not a subject has a brain tumor with a computer visionproblemtoautomaticallycropthebrainfromMRI scans.OthertechniquesusedwereConvolutionalNeural Network,K-MeansClusteringandthehighestAccuracyis given byConvolutionalNeuralNetwork whichisaround 90%. Choudhury et. al. [3] proposed a new CNN-based systemthatcandistinguishbetweendifferentbrainMRI images and label them as tumorous or not. The model's accuracy was 96.08%, and its f-score was 97.3. The model uses a CNN with three layers and only a few preprocessingstepstoyieldresultsin35epochs.Thegoalof this study is to emphasize the significance of predictive therapy and diagnostic machine learning applications. Other techniques used were Support Vector Machine, Convolutional Neural Network, k-Nearest Neighbour, Boosted trees, Random forest and Decision trees To detect, categorise, and segment brain tumours, the suggestedapproachesarecertaintobeveryefficientand accurate.Automaticorsemi-automaticprecision

Thiscanonlybedonewiththeuseofcertaintechniques. For this research, CNN was used to identify and categorise data with the use of a proposed automated segmentationtechnique.Convolutionalneuralnetworks, conditionalrandomfields,supportvectormachines,and genetic algorithms are some of the other methods.CNN hasthebestefficiencyandaccuracy,ataround91%and 92.7%, respectively. In [1] In this study, we use GLCM features and a Multilayer Perceptron neutron to interpret MRI images. With MLP, the network is forwarded with one or more layers between the input and output layers. Segmentation with thresholding, feature vector extraction with GLCM by declaring the four angles-energy, entropy, contrast, and variance, and model learning are all components of the suggested

methodthatmakeuseofthisneuralnetworktechnology. Images that have been filtered or equalised prior to thresholding. Extracting features from data is a kind of datareduction.Theextractedcharacteristicsareusedas training data for a neural classifier. Twenty magnetic resonance imaging (MRI) scans of the head are utilised for testing the suggested ISO algorithm. After the histogram has been equalised, segmentation is used to remove the tumour areas from the whole picture, allowing for a more accurate assessment of the tumor's location within the MRI. The retrieved photos are also usedtomakeacall.

Inthisstudy, weuse theSupportVectorMachine(SVM) technique for detecting brain tumours in MRI data. The Support Vector Machine (SVM) is a statistically-based supervisedlearningsystem.Representingapicturewith DWT. When doing SVM classification, a Simulink model is employed. Using support vector machines (SVM), a prototype is shown in this article that combines fast performance with excellent detection accuracy. To classify tumours, it is necessary to first choose the appropriate pictures for analysis (known as "preprocessing"). Once the tumor's size and form have been identified, the next step is feature extraction. Following this, picture data is used to train a support vector machine. The DICOM format is then used for SVM classification. At last, a diagnosis of tumour is made. Predictive values (PPV) of 81.48 percent and negative predictive values (NPV) of 82 percent are determined. There were22genuinepositives,5genuine negatives,5 genuinepositives,and22genuinenegatives.

In [3] In this study, we use the CNN technique to detect MRI images. Brain tumour identification relies on MRI scans, which are processed to increase precision. Neurons and convolution layers are the basis of a convolutional neural network (CNN). A system's behaviour may be more succinctly represented via clustering, which is used to find naturally occurring groupsinhugedatasets.Thegoalofclusteranalysisisto reveal hidden patterns in large datasets. To locate the areasofmalignancyinimaging,patchextractionisused. CNN's architecture is tailored to exploit the twodimensional nature of an input picture. Segmentation, detection, and extraction are all components of postprocessingforMRIscans.Amongthesystem'sbenefitsis improvedsegmentation.TheaccuracyrateforMRIscans is88%.Applyinganeuralnetworkimprovesaccuracy.

In [4] In this study, the MRI images are identified with the help of a Recurrent Neural Network (RNN). To scale up and scale down the network's nodes, the BP NN activation function was first used. The number of nodes inthehiddenlayerwassetto270andthenbroughtback down to 230 using the log sigmoid function. Finally, we've reached the optimal performance for RNN thanks to a bump in the node count to 300. For optimal

efficiency, we use an Elman network. When the number of nodes is increased, so does the amount of performancemistake.Elmannetworks,whenusedinthe recognition process, were shown to be both quick and accurate compared to other ANN systems. When comparedtoElman's88.14%,ourratiowas76.47%.

III.PROBLEM STATEMENT

Image segmentation and classification has a number of challenges, such as the lack of a universally applicable standardmodel.However,pickingthebestapproachfor each givenCreating a positive image is challenging. So, there is no universally accepted method for picture recognitionandcategorization.It'sstillahugeroadblock for AI vision systems. The method did not take into account the classification of photographs of different clinical disorders, illness types, or disease stages. Because of the high concentration of pure nodes in the system,itisvulnerabletooverfitting.

Developersproposedadeeplearningconcepttoidentify brain tumours automatically using MRI scans, and then analysedtheresultstodeterminehowwellitworked

The Contribution of Proposed Work

 By using a novel boosted adaptive anisotrophic diffusion filter, histogram equalizationthe image get enhanced.

 Segmentation is done in two phases.The first phase is brainsection;thetumourareaisextractedusingahybrid deformable model with a fuzzy approach and a superpixel-basedadaptiveclustering.

 The features are extracted based on texture and tetrolet transform, and the extracted features are combined by using Harish hawks optimization algorithm.

 The suggested technique is intended to distinguish betweennormalbraintumoursandabnormalbrainMRI imagesbytheCNNclassifier

Clustersofimproperlydevelopingbraintissuearecalled tumours, and they may have devastating effects on the centralnervoussystem.

Furthermore, the proliferation of tumour cells might cause abnormal capabilities of the mind. It's also important to note that many different kinds of tumours lead to the gradual enlargement and eventual death of braincells[1].However,thesurvivalrateandnumberof treatmentchoicesforpeoplewithbraintumoursgreatly improves if they are diagnosed at an early stage. While benigntumoursdevelopmoreslowlyandposelessrisk, classifyingthemusingalargenumberofMRIimagesisa

time-consuming and labor-intensive process. / Highqualitymedicalpicturesmaybeobtainedwiththeuseof magnetic resonance imaging (MRI). Medical professionals often use this imaging method to spot problems in the brain. revealing the development of malignancies throughout time. Automatic medical analysis [2] relies heavily on MRI images. They improve the visual representation of the various brain areas by providing anatomical specifics. Researchers have developed a number of ways for identifying and classifying brain tumours using MRI images. There is a wide spectrum of methods, from more traditional medical image processing to cutting-edge machine learningmethods.

To learn from unstructured and unlabeled data without human supervision, deep learning (DL) is a kind of machine learning. In recent years, DL approaches and modelshaveshowneffectiveinsolvingawidevarietyof difficult issues that need a high degree of accuracy and rely on hierarchical feature extraction and data-driven self-learning. Among its many uses, deep learning has been put to work in areas such as pattern recognition, object detection, voice recognition, and decision-making [3]. The largest obstacle facing DL is the massive quantity oftrainingdata required.Asan example,inthe field of healthcare, there is a lack of publicly accessible medical data that may be used to train deep learning models.

The main reason for this is the concern for personal information security. Because of this, the medical industry has relied heavily on transfer learning to make upforthepaucityofdata.Thisisanexampleoftransfer learning, in which a deep learning model that was first trainedforonetaskisthenappliedtoanewissue.Thisis often done in situations when there is insufficient training data [4]. In this research, we apply transfer learningtocreateadeeplearningmodelthatcanidentify and classify brain tumours in MRI scans. All that's needed to construct the proposed model are three pretraineddeeps:

➢ Here, we show off a CNN-based, fully-automatic methodforidentifyingandclassifyingbraintumours.

To further extract deep information from brain MR images,weusepre-trainedmodels.Onadatasetofbrain MRI images with 2 classifications (normal/tumor), we tested 3 pre-trained models and an ensemble of these pre-trainedandCNNmodels.

➢ The present study uses densenet121 and densenet169 with transfer learning on Kaggle datasets to distinguish between benign and malignant brain tumours. Tumor detection, classification, and growth rateestimationareallenhancedbyMRI

A. Data Pre-processing step

Preparing Data Training dataset size is very important fordeepconvolutionalneuralnetworks.Thefirstversion of the model was built using Keras TensorFlow's Image DataGenerator.

In order to train the proposed system on a sufficient numberofMRIpictures,theimagedatasetisaugmented with random modifications (rotations, height and breadth shifts, brightness changes, etc.). The recommended classifier will never be exposed to the same picture again because to the data augmentation parametersthatwerecarefullyselected.

B. Data augmentation

There is a strong correlation between the amount of datasets utilised during training and the final performance of deep convolutional neural networks. To better train the proposed system on MRI pictures, we augment the original dataset with additional MRI images. (pivots, scale adjustments, luminance) TensorFlow is modified with the use of the ImageDataGenerator function in Keras. What was done to improve the data so that the suggested classifier can makereliableidentifications.Youwillneverseearepeat of yourself. This process improves the model's generalisationabilities.

C. Rezing and Crope

The removal of the brain from the picture backdrop is the initial stage in this process [19]. The approach showed here use OpenCV to pinpoint a bounding box's extrema.It'sworthnotingthattheMRIimagesutilisedin this research come in a range of sizes according to their geographical origins. To ensure uniformity, the photos areresizedto64x64x1..

D. Data Spiliting In this analysis, we partition the data into smaller subgroups. The suggested approach will be tested, validated, and refined over the course of three phases. a case study in comprehensive education. The modelcanbefitusingthefirstsubset.About80%ofthe whole dataset is included here. The remaining portion hasvanished.Thesystemwillbetestedandvalidatedat thesamerateof80%each.

E. Convolutional Neural Network CNN is well-known because it has gotten better at classifying images. CNN uses the data that is given to it to automatically collect features. It is a well-known DL architecture with feedforward connections between each layer. Deep architecture helps these networks learn complex functions that a simple neural network can't learn [20]. CNN is the brains behind computer vision, and it can be used to classify objects, monitor activities, and create

medical images. Its preprocessing is simple and small compared to other neural classifiers because it has an internal filter. A typical CNN architecture includes the followingcomponents:

(i)Convolution

(ii),pooling

(iii),activation

(iv),andadenselayer

(v)areallstepsintheclassificationprocess.

F. Transfer Learning Transfer learning (TL) is a deep learning strategy that marries an existing model trained on a large dataset with a new model trained on a different dataset to tacklethesameproblem. In general, CNN performs better with more data than with less. In CNN situations when the datasets are small, TL can be useful. The field of TL has expanded in recent years, findingapplicationsinobjectdetection,medicalimaging, and picture categorization [21] Large Datasets, such as ImageNet, were utilized to train the models so that they can extract useful characteristics. Applications utilizing smaller data sets, such as brain MRI data, are common. One of the benefits of TL is that it shortens the time it takes to complete training. Procedures, avoiding snug fits, training with fewer data, and motivating improved performance. CNN model training is complete. Efficientb0, DenseNet121, and DenseNet169 were utilizedinourstudy

Conclusion:

In terms of classification, dense convolutional layers neural pathways (CNNs) have been greatly explored in the science and new firms. In this research paper, evaluate the effectiveness of a deep neural network model for a classification algorithm linked to the diagnosis of brain cancer. The extension applied to the ResNet model shows that the deep learning algorithm used in natural image processing can provide superior efficiency in the analysis of medical data. The proposed device was designed to identify the brain tumour of the Magnetic Resonance Imaging (MRI) brain. There are many phases novel boosted adaptive anisotrophic diffusion filter is used for noise removal, to detect the tumor part from the brain image. Segmentation extracts thetumorportion.Textural featureextractionisusedin proposed system. Classification uses CNN classifiers; proposed system achieved 98.3% of accuracy. Future enhancement is to perform with the Perfusion based MRIimageswhichisquitecomplex.

REFERENCES:

1. Saba, T., Mohamed, A.S., El-Affendi, M., Amin, J., & Sharif,M.(2020).Braintumordetectionusingfusion ofhandcraftedanddeeplearningfeatures.Cognitive SystemsResearch,59,221-230.

2. Sharif, M., Amin, J., Nisar, M.W., Anjum, M.A., Muhammad,N.,&Shad,S.A.(2020).Aunifiedpatch based method for brain tumor detection using features fusion. Cognitive Systems Research, 59, 273-286.

3. Toğaçar, M., Ergen, B., &Cömert, Z. (2020). BrainMRNet: Brain tumor detection using magnetic resonance imageswitha novel convolutional neural networkmodel.MedicalHypotheses,134,109531.

4. Veeramuthu,A.,Meenakshi,S.,&Darsini,V.P.(2015). Brain image classification using learning machine approach and brain structure analysis. Procedia Computer Science, 50, 388-394. www.sciencedirect.com

5. Kumar, S., Dabas, C., &Godara, S. (2017). Classification of brain MRI tumor images: A hybrid approach.Procediacomputerscience,122,510-517. www.sciencedirect.com

6. Jana, G.C., Swetapadma, A., &Pattnaik, P.K. (2018). Enhancing the performance of motor imagery classification to design a robust brain computer interface using feed forward back-propagation neuralnetwork.AinShamsEngineeringJournal,9(4), 2871-2878.www.sciencedirect.com

7. Sidhu, H.S., Benigno, S., Ganeshan, B., Dikaios, N., Johnston, E.W., Allen, C., & Taylor, S.A. (2017). Textural analysis of multiparametric MRI detects transitionzoneprostatecancer.Europeanradiology, 27(6),2348-2358.

8. Anaraki,A.K.,Ayati,M.,&Kazemi,F.(2019).Magnetic resonance imagingbased brain tumor grades classification and grading via convolutional neural networks and genetic algorithms. Biocybernetics andBiomedicalEngineering,39(1),63-74.

9. Fabelo, H., Ortega, S., Lazcano, R., Madroñal, D., M Callicó, G., Juárez, E., &Piñeiro, J.F. (2018). An intraoperative visualization system using hyperspectral imaging to aid in brain tumor delineation.Sensors,18(2),430.

10. Zhang,J.,Shen, X.,Zhuo,T., &Zhou,H.(2017).Brain tumor segmentation based on refined fully convolutional neural networks with a hierarchical diceloss.arXivpreprintarXiv:1712.09093

11. Wang,N.,Shang,Y.,Chen,Y.,Yang,M.,Zhang,Q.,Liu, Y., &Gui, Z. (2018). A hybrid model for image denoising combining modified isotropic diffusion model and modified Perona-Malik model. IEEE Access,6,33568-33582.

12. Chang, Y., Jung, C., Ke, P., Song, H., & Hwang, J. (2018). Automatic contrast-limited adaptive histogram equalization with dual gamma correction.IEEEAccess,6,11782-11792.

13. Haleem, M.S., Han, L., van Hemert, J., Li, B., Fleming, A., Pasquale, L.R., & Song, B.J. (2018). A novel adaptivedeformablemodelforautomatedopticdisc and cup segmentation to aid glaucoma diagnosis. Journalofmedicalsystems,42(1),20.

14. Wang, Y., Karimi, H.R., Shen, H., Fang, Z., & Liu, M. (2018). Fuzzy-modelbased sliding mode control of nonlinear descriptor systems. IEEE transactions on cybernetics,49(9),3409-3419.

15. Uziel, R., Ronen, M., &Freifeld, O. (2019). Bayesian Adaptive Superpixel Segmentation. In Proceedings of the IEEE International Conference on Computer Vision,8470-8479.

16. Pradhan, J., Kumar, S., Pal, A.K., & Banka, H. (2018). A hierarchical CBIR framework using adaptive tetrolet transform and novel histograms from color and shape features. Digital Signal Processing, 82, 258-281.

17. Yang,F.,Dogan,N.,Stoyanova,R.,&Ford,J.C.(2018). Evaluation of radiomic texture feature error due to MRI acquisition and reconstruction: a simulation study utilizing ground truth. PhysicaMedica, 50, 2636.

18. Sihwail, R., Omar, K., Ariffin, K.A.Z., &Tubishat, M. (2020). Improved harrishawks optimization using elite opposition-based learning and novel search mechanism for feature selection. IEEE Access, 8, 121127-121145.

19. Chuang, T.Y., Han, J.Y., Jhan, D.J., & Yang, M.D. (2020). Geometric Recognition of Moving Objects in Monocular Rotating Imagery Using Faster R-CNN. RemoteSensing,12(12),1908.

20. Ćiprijanović, A., Snyder, G.F., Nord, B., & Peek, J.E. (2020). DeepMerge: Classifying high-redshift merging galaxies with deep neural networks. AstronomyandComputing,32,100390.

21. Sarhan, A.M. (2020). Detection and Classification of Brain Tumor in MRI Images Using Wavelet Transform and Convolutional Neural Network.

22. Singh, S.P., Wang, L., Gupta, S., Goli, H., Padmanabhan, P., &Gulyás, B. (2020). 3D Deep Learning on Medical Images: A Review. arXiv preprintarXiv:2004.00218.

23. Chaubey, N.K., &Jayanthi, P. (2020). Disease Diagnosis and Treatment Using Deep Learning Algorithms for the Healthcare System. In Applications of Deep Learning and Big IoT on PersonalizedHealthcareServices,IGIGlobal,99-114.

24. Joseph, T., Kalaiselvan, S.A., Aswathy, S.U., Radhakrishnan, R., &Shamna, A. R. (2020). A multimodal biometric authentication scheme based on feature fusion for improving security in cloud environment. Journal of Ambient Intelligence and HumanizedComputing,1-9.