Brain Tumor Detection and Classification Using Intelligence Techniques: An Overview

A tumor is carried on by rapid and uncontrolled cell growth in the brain. If it is not treated in the initial phases, it could prove fatal. Despite numerous significant efforts and encouraging outcomes, accurate segmentation and classification continue to be a challenge. Detection of brain tumors is significantly complicated by the distinctions in tumor position, structure, and proportions. The main disinterest of this study stays to offer investigators, comprehensive literature on Magnetic Resonance (MR) imaging’s ability to identify brain tumors. Using computational intelligence and statistical image processing techniques, this research paper proposed several ways to detect brain cancer and tumors. This study also shows an assessment matrix for a specific system using particular systems and dataset types. This paper also explains the morphology of brain tumors, accessible data sets, augmentation methods, component extraction, and categorization among Deep Learning (DL), Transfer Learning (TL), and Machine Learning (ML) models. Finally, our study compiles all relevant material for the identification of understanding tumors, including their benefits, drawbacks, advancements, and upcoming trends.


I. INTRODUCTION
An unchecked expansion of brain tissues is known as a brain tumor. It produces pressure in the skull and interferes with the brain's natural functioning. Brain tumor comes in two different types: Benign (non-cancerous) and Malignant (cancerous). Among them, malignant tumors grow quickly in the brain, damage the normal tissues, and may replicate themselves in other parts of the body [1], [2], [3]. Brain tumors are graded into four different categories: Grade I: These tumors do not spread quickly and develop slowly. These are connected to a higher chance of enhanced order and may be surgically eliminated nearly entirely. One such tumor is a pilocytic astrocytoma.
Grade II: Although they may migrate to surrounding tissues and advance to higher grades, these tumors also grow The associate editor coordinating the review of this manuscript and approving it for publication was Wei Wei . over time. These tumors may detect even though treatment is taken by the patient. An oligodendroglioma tumor is an example of an overtime growth tumor.
Grade III: The growth of these tumors has been quicker than grade II malignancies and could spread to adjoining tissues. Such tumors require post-operative chemo or radiotherapy because surgery alone would be insufficient to treat them. Aden squamous astrocytoma is an indication of such a tumor. Grade IV: The most dangerous and likely to spread malignant tumors are in this category. They might even use blood vessels to speed up their growth. An illustration of one of these tumors is glioblastoma multiforme [3], [4], [5].
Brain tumors must be identified in time and appropriately be classified in order to get proper treatment and endure for patients. Because of the several vulnerabilities including different shapes, sizes of tumors, appearance, positions, scanning parameters, and modalities detection of brain tumors is a very challenging job to perform [5]. To attain this task a number of traditional and intelligence techniques are being used. Typically, traditional approaches like Leksell Gamma Knife, Gamma Knife (GK), and Radioactive beams are helpful in diagnosing the lesions, but this process includes human involvement and is often a time-consuming task to perform [6]. For brain tumor identification, many medical imaging modalities like Computer Tomography (CT), Magnetic resonance imaging (MRI) scans, and Poisson Emission Tomography (PET) are employed. Also, A unique MR technique called chemical exchange saturation transfer (CEST) makes it possible in imaging some substances at concentrations that are too low to affect the contrast of conventional MR imaging and too low to be directly identified in MRS at usual water imaging resolution. Among them, MRI scan is a non-invasive method that shows the internal body structure with the help of magnetization and microwave pulses. For brain tumor diagnosis, three categories of magnetic resonance image patterns are used: Fluid Attenuated Inversion Recovery (FLAIR), T1 weighted, and T2 weighted. The problem of identifying and detecting tumor-infected areas using brain MRI is crucial [6]. The human visual system has a minimal ability to notice tiny variations brought on by the Magnetic Resonance Image's increased complexity (MRI). Recently, a number of investigators developed Systems for computer-aided diagnosis (CAD) to help radiologists make precise diagnoses [6]. Although Leksell Gamma Knife is a better approach to diagnosing tumors, because of the presence of necrosis in the brain the finding suffers. Therefore, effective machine learning should be adopted in order to solve this problem. Authors in [7] have proposed a novel method with the amalgamation of a Random Forest classifier along with a voxel clustering algorithm. Similarly, conventional diagnosis processes including Leksell Gamma Knife are time-consuming processes, therefore authors in [8] have introduced a semiautomated method using an unsupervised FCM clustering algorithm for accurately segmenting the lesion volume. A pipeline of four algorithms namely K-means, Fuzzy K-means and Gaussian Mixture Model (GMM), and Gaussian Hidden Markov Random Field (GHMRF) has been proposed for the segmentation of brain tumors by the authors in [9]. Authors in [10] propose a two-stage mechanism for the assessment in dose escalation and eliminate the need for multispectral MRI data to analyze the image. The proposed framework incorporates the FCM algorithm in defining a novel method named a fully automatic method for necrosis extraction (NeXt). Although ML approaches are quite efficient in handling the MRI images for accurate detection of the tumor region, with the availability of complex, large volumes of data and high computing devices, deep learning models are being cast-off for achieving advanced performance. Therefore, to understand the detailed learning mechanism of these intelligence techniques the proposed work is aimed in presenting: 1) The proposed work incorporates various deep learning and machine learning mechanisms adopted for the detection and classification of brain tumors from MRI images. Study was carried out for about 100 articles collected  from various sources like ScienceDirect, Springer, IEEE, etc. 3) A separate analysis of both approaches has been carried out and various findings are tabularized individually.

2)
4) Further, a research gap analysis has been carried out to differentiate between the importance of DL over ML. 5) Various findings like datasets, deep models, classification approaches, parameters, future research directions along with the importance of using 3D models and attention-based mechanisms are being discussed followed by our proposed work.
Organization of manuscript: introduction in section I is followed by a literature review given in section II. It includes the various studies categorized among DL and ML in a separate section. A research gap analysis of the underlying technologies is also been given in this section. Further, the various findings are given in Section III. Section IV demonstrates our proposed work. The manuscript ends with a conclusion followed by references.

II. LITERATURE REVIEW A. DEEP LEARNING TECHNIQUES
In recent years, a lot of research has been directed toward the adaptation of deep learning models in diagnosing brain tumors. Academicians have put in their efforts and with the help of high-end computing devices, higher accuracy has been achieved. Convolutional neural networks (CNN), which include input, output, hidden layers, and hyperparameters, are often called Deep Learning (DL) [5]. It uses supervised classification and generates feature maps by having the kernel convolve all around the input image. Automatic-based feature extraction is both possible with DL models. Apart from its usefulness for medical condition detection, it has some shortcomings, including the requirements to design complex models, fine-tuning of hyper-parameters, the requirement of large data set, and time and effort to training/testing. As per recent research, significant data augmentation methods like resizing, rotation, scaling, and transformation are enforced to tackle the big data availability problem. A trained NN is used in transfer learning techniques to extract similar properties from an application-specific dataset [1]. For brain tumor identification current TL methods like RESNET-100, VGGNET, Google-Net, AlexNet, etc. are applied. The various deeplearning techniques used by the researchers in the past are summarized in Table 1.
With the recent developments in technology, 3D scanning is also being used for the analysis of tumors. 3D image processing for brain tumor detection and classification has been described in [48]. It used various deep learning frameworks, such as MobileNetV2, MobileNetV3 small, MobileNetV3 big, VGG16, VGG19, and custom CNN models. CNN achieved the highest accuracy. It offers a solution that combines a CNN built with Keras and Tensor flow with a fully-featured cross-platform application built with   VOLUME 11, 2023 PyQt5 and MariaDB, all of which are designed for usage in medical settings like hospitals and clinical images. The primary goal of this work is to characterize a brain damaged by a tumor using real-world data and identify abnormal pixels [51]. In [52] authors describe a pre-processing, data augmentation, segmentation, and binary classification of brain tumors implemented with a 3D medical image. In this context, classification is performed using two distinct classifiers: Dense-Net and Dark-Net. On the BRATS 2018 dataset of 3D-MR images, the suggested framework obtained an accuracy of 98.67% and a dice similarity coefficient (DSC) of 97.91% for segmentation. For brain tumor classification on 3D-MR images from the BRATS 2018 dataset, the suggested framework obtained a DSC of 98.14%, an accuracy of 98.26% using the Dense-Net classifier, and a DSC of 96.4%, an accuracy of 96.52% using the Dark-Net classifier. A higher level of accuracy was achieved by the Dense-Net classifier compared to the Dark-Net classifier. In addition, they have compared this framework to earlier research, and the results show that custom CNN obtains higher classification accuracy. Some of the 3D-based methods are given be Table 2.

B. MACHINE LEARNING METHODS
Pre-processing, segmentation, extraction of features, and categorization are the four key phases of ML techniques used to diagnose brain tumors.

1) PREPARATION
To produce accurate diagnoses in the clinical field, precise imaging is essential. The efficiency of clinical images is influenced by the artifact acquisition methods, like magnetic resonance scans, CT, and PET. A Magnetic resonance scan's real images could contain a lot of unwanted and pointless details. Magnetic resonance imaging is impacted by Rician noise [36]. It is challenging to remove Rician distortion since it is signal-sensitive. Pre-processing techniques including filtration, intensity correction, and skull stripping is being used to maintain the original visual characteristics.

2) SEGMENTATION
It is a technique used to obtain areas of interest from digital images. The tumor's position must be distinguished from the MR brain scans, which is crucial. For segmentation, numerous supervised methods are available, including thresholding, soft computing technique, atlas-based, Neural Networks (NNs), clustering, etc. Thresholding methods include global, adaptable, Otsu's, and histogram-dependent techniques. There are two unsupervised clustering methods namely K-means clustering and fuzzy C-means clustering. It successfully separates brain MRI scans into Gray Matter (GM), Cerebrospinal Fluid (CSF) as well as White Matter (WM). Segmentation techniques that draw inspiration from nature include Particle Swarm Optimization (PSO) and Genetic Algorithm. Recent studies show that DL frameworks like Convolutional Neural Networks (CNN), Mask-Recurrent Neural Networks, and UNET outperform conventional methods in segmentation.

3) FEATURE EXTRACTION
While extracting features, properties of brain MR scans such as shape, structure, wavelet, and Gabor are retrieved. The Gray-Level Co-occurrence Matrix (GLCM) is commonly studied. A second-order statistical method is used to evaluate textural features like energy, correlation, and intensity. Wavelet data is derived using the Discrete Wavelet Transform (DWT). The approximation coefficients are obtained and it is applied to an original image, and then the feature vector is selected. Both automatic features produced by DL techniques like Convolutional Neural Networks, ResNet, Capsule Networks, and handwritten features have shown success. To decrease the number of features, PCA and Genetic Algorithms are utilized.

4) CLASSIFICATION
Benign and malignant tumors are the most prevalent forms of brain tumors. The three types of malignant tumors include hypothalamic, gliomas, and malignant tumors. Table 3 shows a summary of some ML methods.

C. GAP ANALYSIS OF ML AND DL METHODS
Further, to analyze the research gap between the existing machine learning and deep learning approaches, the study has been directed towards summarizing the various literature work incorporating both technologies which are presented in Table 4. This table compromises the details with respect to methodology, algorithms, gap analysis, and dataset used by the authors.

III. FINDINGS
Expert radiologists do brain tumor segmentation and classification. ML and DL may help radiologists to make better decisions. This paper summarizes current strategies for automated brain tumor categorization. Histogram equalization, median, Gaussian, and Wiener filters preprocess MRI images. There are six forms of segmentation: clustering, statistical, CNN, region, and threshold-based [37]. K-means Researchers often utilize C-means clustering and adaptive global thresholding. Deep learning-based segmentation allows for more precise tumor extraction [26]. GLCM and DWT largely extract features. GLCM returns texture characteristics, whereas DWT returns approximation coefficients. Deep learning architectures automate feature extraction. ResNet [4], [11]. PCA and bio-inspired algorithms like PSO are used to reduce dimensionality. Choosing the optimal characteristics for categorization is challenging. Hence a hybrid technique integrating several features is utilized. ML and DL techniques are used to classify data. multi-kernel SVM Binary classification uses linear, RBF, and Cubic. These findings are comparable to VGG19 and ResNet. ANFIS, a fuzzy-ANN hybrid, performs better for binary classification. However, the database does not capture all tumor forms and grades. Or they have to obtain MRIs from nearby hospitals. As a result, comparing the performance of various approaches is difficult. A common database of all tumor kinds is required for future study. The deep learning approach can extract more detailed features from the dataset for segmentation and classification. Transfer learning techniques provide better prediction results for deep learning approaches in the effective detection of brain tumors. The machine learning approach gives better performance when the dataset is small, whereas, with large datasets deep learning, models are efficient. The deep learning approach used several pre-processing techniques like scaling and normalization to enhance desired features. Preprocessing techniques of machine learning including filtration, intensity correction, and skull stripping are being used to maintain the original visual characteristics with a limited data set. The primary drawback of machine learning technology is that it is complicated, with a large number of parameters increasing with the execution time and system requirements for implementation. The deep learning approach offers low complexity, where the features are self-learned by the network. Figure 1 and Figure 2 demonstrate the valuation of Machine Learning and Deep Learning methods over accuracy.
Further, the various findings from the literature are being discussed in separate categories like datasets used, different tumor classification approaches, deep learning models, parameters used, limitations of existing approaches, and future research directions.

A. DATASETS USED
The researchers make use of a variety of datasets that are available to the general public in order to test the proposed  methodologies. In this section, we will go through various challenging datasets that are both significant and crucial. The BRATS datasets are considered to be the most difficult MRI datasets [73], [74], [75]. BRATS Challenge is issued at different times throughout the years, and more recent challenges have had a resolution of 1 mm3 voxel [76]. Employed two benchmark datasets and one dataset obtained from qualified radiologists. These datasets include 15 photographs of patients, and each patient included 9 slices of imaging data. The core dataset that was used was known as the digital imaging and communication in medicine (DICOM) dataset. Twenty-two photos from the DICOM collection, some of which depict tumor-infected brain tissue, have been taken into consideration for the purpose of this investigation. This dataset did not contain any images that represented the ground truth. The brain web dataset [77] was used as a supplemental source of information for this study. The whole threedimensional simulated brain MR data that is included in this paper were obtained using three modalities: proton densityweighted MRI, T1, and T2-weighted MRI, and T1-weighted MRI. The BRATS 2017 dataset [78] was utilized by Shubhashis Banerjee and Francesco Masulli. This dataset comprises data from the BRATS 2012/13/14, and 2015. A total of 210 HGG cases and 75 LGG instances of brain tumors are included in the dataset. The patient's MRI scan includes four distinct MRI sequences: the initial (T1) sequence, the T1 & T2 weighted sequence, and the Fluid Attenuated Inversion Recovery (FLAIR) volume with 155 two-dimensional slices at a resolution of 240 by 240. The BRATS training dataset, which consists of 274 multi-modality MRI images of people with gliomas, is used by the researchers Ali Isin, Cem Direkoglu, and Melike sah (both high and low grades). For the purposes of testing, a total of 110 scans were taken from ground truths and unknown grades.

B. TUMOR CLASSIFICATION APPROACHES
The input data is sorted using classification techniques into a variety of separate classes., after which training and validation are carried out using both known and unknown instances. The classification of tumors into relevant classifications is a widespread application of machine learning, tumor as well as non-tumor, and malignant and benign tumors. Supervised methods include KNN, support vector machine, nearest subspace classification model, and representation classification model. Fuzzy C Means, hidden Markova random field, and self-organization map, are examples of unsupervised approaches [67], [68], [69], [70], [71].

C. DL MODELS
Deep learning (DL) models, as opposed to shallow Machine Learning (ML) techniques, are founded on the principles of learning data representations as well as learning hierarchical features. Deep learning techniques are used to categorize brain tumors, and these techniques find the descriptive data that most properly describes the many forms of brain tumors. The classification of brain tumors shifts away from being driven by manually created characteristics and toward being driven by data due to the nature of deep learning [87]. In the domain of deep learning technics, a convolutional neural network is one of the most popularly utilized ones for the categorization of brain tumors, and a significant amount of progress has been made [88]. There are a few different  approaches that may be taken when classifying brain tumors, which can be seen in the research that was looked through. The difference includes the following aspects: (i) the dataset that was used for categorization, which included the types of tumors; (ii) the pre-processing as well as data augmentation methods that were incorporated; (iii) The use of ROI segmentation as a preliminary step in classification; The use of either a pre-trained or custom-designed deep learning technique; and (iv) the ROI segmentation question. For example, Bada and Barjaktarovi'c [88] used contrast-enhanced T1-weighted brain tumor MRI images that were readily available to the public [89]. Meningioma, glioma, and pituitary tumor scans are included in the collection, as well as images from the three anatomical perspectives of axial, sagittal, and coronal. The images were preprocessed using several techniques, such as scaling and normalization, among others. To increase the size of the training dataset, the photographs in the dataset are also flipped vertically and rotated via a 90-degree angle. Additionally, they employed a specially created CNN classifier trained with the Adam optimizer using a mini-batch size of 16 and then tested the Classifier with 10-fold cross-validation. A Glorot initializer is used to get the convolution layers' weights started off in the right direction. The measures utilized to assess the model's performance were the highly sensitive, selectivity, accuracy, recall, and F1-score. Meningiomas [49], [50], [51], [52], [53], [54], [55], [56], [57], [58], [59], [60], [61], [62].

D. PARAMETERS
MRI images are identified and categorized using an automated technique, as described in [31]. The Super Pixel Methodology forms the foundation for this strategy, as does the categorization of every Super-the pixel. When attempting to categorize each superpixel as either tumorous or normal; the extremely randomized trees classification model is evaluated alongside the support vector machine. This methodology utilizes two datasets, which are referred to as a dataset of BRATS released in 2012 and 19 MRI FLAIR images respectively. The findings indicate that the utilization of the ERT classifier yields satisfactory results for this strategy. An instinctive organization technique is employed to recognize a tumor using a convolutional neural network with 3 × 3 tiny kernels [84]. The tumor is ordered using this technique. The method won the BRATS Challenge in 2013 by simultaneously placing first in the whole, core, and improving areas of the dice similarity, coefficient metric (0.880, 0.830, and 0.770). In the [32], the Alexnet model convolutional neural network is rummage-sale to concurrently identify multiple sclerosis and normal tumors. A convolutional neural network was successful in correctly categorizing 98.67% of the photos into one of three categories. To segment brain cancers from MRI scans, a multi-stage Clustering framework was proposed in [54]. According to [85], there is a method for categorization and segmentation that makes use of CNNs that is both efficient and effective. Image-Net was utilized in the suggested method in order to extract features. According to the findings, the classification was accurate to the extent of 97.5%, while the segmentation was accurate to the extent of 84%. In the study referred to as [86], Analysis of multiphase MRI images for tumor grading has been conducted, and the results of base neural networks and deep learning structures have been contrasted and compared. According to the findings, the performance of the network which is measured by the specificity and sensitivity of CNN has increased by 18 percent when related to the efficiency of neural networks. In the paper [33], the authors present a deep learning-based supervised technique for detection variations in artificial opening radar scans. This technique provides a dataset that had an adequate amount of data volume and variety for the purpose of training the DBN with the help of the input photographs and the images that were acquired by applying structural operatives on those images. The finding accuracy of this technique demonstrates the applicability of techniques based on deep learning for the purpose of finding solutions to change detection challenges.

E. LIMITATIONS OF ML OVER DL APPROACHES
Recent research on the diagnosis of brain tumors is examined in this survey; the findings suggest that there is an opportunity for further development in this area. Noise is introduced into an MRI scan during the image capture process, and removing this noise is a complex process [20], [21], [22], [23]. Due to the tentacles and dispersed features that are characteristic of brain tumors [23], [24], [25], accurate segmentation is a thought-provoking task. In order to achieve better categorization, one of the most significant tasks is to select and retrieve the optimal features, as well as determine the right amount of training and testing samples [26], [27]. The fact that deep learning models can autonomously learn new features is one of the reasons they are getting popular. On the other hand, these models need a portion of memory and a lot of dispensation control. Therefore, it is still necessary to develop a lightweight computing framework that can produce a high ACC in a shorter length of time. The following is a list of the primary difficulties associated with detecting brain tumors. The glioma tumor and the stroke tumor do not contrast very well with one another. It is made up of tentacles and scattered components, both of which make the segmentation and categorization procedures far more difficult [83]. The identification of a tiny volume of the tumor remains difficult since it is possible for it to be recognized as a normal area [29]. Some of the existing approaches perform admirably for a complete tumor region but not for other regions (whether enhanced or not), and conversely [90], [91], [92], [93], [94], [95].

F. FUTURE RESEARCH DIRECTIONS
This survey includes all of the significant features as well as the most recent work that has been done along with their constraints and obstacles. The researchers will benefit from gaining a better understanding of how to do new research in the appropriate manner within a reasonable amount of time. Even though deep learning approaches have made substantial contributions, there is still a need for a generic approach. These methodologies produce better outcomes once training and testing are carried out on achievement features (intensity range as well as resolution) that are comparable; moreover, the robustness of the methodologies are directly impacted by even the slightest change between the training imaginings and the testing imaginings. In the future, studies may be conducted to detect brain cancers more precisely, with actual patient information since somewhat average (various image capture methods) (scanners). Combining handcrafted characteristics with deep features has the potential to enhance classification accuracy. Similar to this, lightweight technologies like quantum machine learning are crucial in enhancing accuracy and efficacy, which in turn cuts down on the time required by radiologists and raises the percentage of patients who survive their illnesses. An attention-based mechanism improves brain tumor segmentation outcomes and reduces computational complexity issues. To be more precise, an image processing and attention mechanism are used to extract the desired area of the image, and then a pre-trained encoder part extracts the fewest but most crucial features to further improve the efficiency of the results. One of the most studied ideas in the field of deep learning is attention, which is used to solve issues like neural machine translation and image captioning. The attention mechanism idea is supported by a number of theories, including Seq2Seq models, encoders, decoders, hidden states, context vectors, and others. Channel attention, Spatial attention, and Block attention are some of the useful methods. Some of the suggestions and possible improvements made by the published review articles includes: Further practice of hybrid-based learning technique is important to obtain strong CAD system [88], Noise estimation is challenging in machine learning and in deep learning lack of interpretability [89], How effectively automatic methods can manage the impact of treatment effects is still being researched [90], Technical issues stemming from the difficulty in defining exactly what deep learning is due to the lack of mathematical and theoretical foundations for many of its core models and techniques [91], Research should carefully consider how to lessen or compensate for observer, spectrum, and selection biases, as well as how to increase reporting transparency [92], Research should focus on optimization technique which will decide number of layers and filters in the model [93], Semi supervised training gives weak performance [94], Absences of transfer learning mechanism leads to weak generalization ability [95], Deficiency of training data and no resolution gives poor performance of CNN [96], With large volume of data quality of image segmentation needed to improved [97], Transfer learning model is required incapacitating overfitting of image [98], Accurate analysis is difficult for vast number of images [99], and Computation is difficult with multiple task [100].

IV. PROPOSED FUTURE WORK
The flowchart for the proposed work is given in Figure 3. This describes the execution of the proposed system in the detection of various diseases using CNN. The entire architecture depicts how the system deals with the recognition and detection of the test image, and below we explain the process of execution. The purpose of this research is to combine feature selection approaches with machine learning to identify pre-illnesses. For the early diagnosis of early diseases in MRI, CT scan, and X-ray images, this system makes use of deep learning techniques and image processing technology. to make feature extraction more efficient, the dataset including defective images from several categories was pre-processed and segmented.
Image Acquisition: In image acquisition, heterogeneous images of the medical dataset collected which contains abnormal and normal samples are gathered from a variety of individuals and converted into image format using a camera or some synthetic dataset.
Pre-processing: There may be difficulties like noise, image blurring, and other concerns since the input data samples were gathered from a range of people. As a consequence, preprocessing methods are used for images in order to reduce noise and improve image quality using modern techniques.
Processing the image is tough due to the fact that it is originally in RGB color format. The RGB to greyscale conversion is required to reduce the complexity of a 3D pixel value to a 1D value. Many applications, such as edge detection, do not benefit from the use of three-dimensional pixels.
Feature Selection: In image processing and data mining, feature selection is critical. It calculates the best subset of predicted characteristics from the original data. A subset of the original characteristics is chosen that retains enough information to distinguish successfully across classes. For feature selection, many search techniques can be utilized such as IG, PCA, and RAE.
Feature Extraction: There are six separate sets of photos taken, from various available datasets. The obtained images are then subjected to image processing methods in order to identify valuable information for future study. Because the gathered photos are of various sizes, it is necessary to transform them to a consistent size for effective preprocessing. The RGB photos are first scaled and transformed to Hue Saturation Intensity (HSI) format. Color perception is greatly aided by the use of HSI color space representation. Masking is then used to eliminate the pixels. Setting the pixel value of a picture to zero or another background value is known as masking. The diseased section of the original picture is then segmented using the K-means segmentation technique. The goal of segmentation is to transform a picture's representation into a meaningful image that is simpler to explore. The best characteristics from this dataset are then selected for accurate categorization via feature selection. Relief-f Attribute Evaluator (RAE), Principal Component Analysis (PCA), as well as Information Gain (IG), are the three techniques of feature selection used in this study.
Classification and Recognition: Disease classification is the process of recognizing a test sample and giving it the appropriate class label. The result of the feature extraction module is to feed the classifier as an input. The classifier will identify the right class label for the input image based on the retrieved characteristics. There are a variety of methods for categorization. Deep learning is one of them. Deep learning employs a variety of artificial neural networks, including CNN, ANN, RNN, and others. The image is sent into CNN, which extracts the most important characteristics as distinct layers. The key benefit of the convolutional neural network is that it lowers the amount of work required by humans to extract characteristics. Implementation Process 1) The images from the dataset can be pre-processed in this phase to prepare the data for subsequent processing. Initially, photos are processed to reduce noise. The photos were then transformed to grayscale and scaled to proper pixels while keeping the aspect ratio constant.
2) The aspect ratio of the picture, the amount of lateral and vertical lines in the picture, the location and number of loops and curves, and other geometrical elements can all be extracted from each image sample through processing. These traits were then combined with the pixel-based information from the image to yield accurate classification outcomes.
3) The input values of a traditional neural network are modified by passing through a sequence of hidden layers. A group of neurons makes up each layer, each of which is totally linked to every neuron in the layer preceding it. The superior performance of CNNs is due to the fact that these networks capture the fundamental features of pictures. This important property of CNN gives the confidence to apply it in the suggested dataset analysis. 4) Download the dataset from open-source websites such as the Kaggle dataset or any unseen dataset. 5) Extract various features by using the CNN model, train module, and save features in a. pkl file. There are below layered concepts in Convolutional Neural Networks: 6) In the dense layer, it will classify the validation image set and show classification results accordingly.

V. CONCLUSION
CAD systems for the detection of brain tumors are developed using brain MRI scans and digital image processing methods like pre-processing, separation, and classification. The classic deep and machine-learning techniques for brain tumor identification are discussed in this work. Various research publications from reputable journals and conferences have been examined, with a full analysis of each work offered. This section provides a summary of commonly used MRI datasets. Although several ML and deep learning methods are used for classification, CNN has shown to be quite accurate. CNN is often used to categorize brain tumors into two types: normal and pathological. The development of an autonomous brain tumor detection system must consider reliability, accuracy, and calculation time. This review examines current methodologies and can be utilized in the future to build effective diagnostic tools for additional brain illnesses that as Alzheimer's disease, Parkinson's disease, dementia, and stroke using various MRI imaging modalities. Implementing this system in collaboration with multiple deep learning algorithms as deep hybrid learning for brain tumor detection and classification will be future work for this study.