Brain tumor detection and classification using machine learning: a comprehensive survey

Brain tumor occurs owing to uncontrolled and rapid growth of cells. If not treated at an initial phase, it may lead to death. Despite many significant efforts and promising outcomes in this domain, accurate segmentation and classification remain a challenging task. A major challenge for brain tumor detection arises from the variations in tumor location, shape, and size. The objective of this survey is to deliver a comprehensive literature on brain tumor detection through magnetic resonance imaging to help the researchers. This survey covered the anatomy of brain tumors, publicly available datasets, enhancement techniques, segmentation, feature extraction, classification, and deep learning, transfer learning and quantum machine learning for brain tumors analysis. Finally, this survey provides all important literature for the detection of brain tumors with their advantages, limitations, developments, and future trends.


Introduction
The central nervous system disseminates sensory information and its corresponding actions throughout the body [1-3]. The brain, along with the spinal cord, assists in this dissemination. The brain's anatomy [4] contains three main parts; brain stem, cerebrum, and cerebellum. The weight of a normal human brain is approximately 1.2-1.4 K, with a volume of 1260 cm 3 (male brain) and 1130 cm 3 (female brain) [5]. The frontal lobe of brain assists in problem-solving, motor control, and judgments. The parietal lobe manages body position. The temporal lobe controls memory and hearing functions, and occipital lobe supervises the brain's visual processing activities. The outer part of cerebrum is known as cerebral cortex, and is a greyish material; it is composed of cortical neurons [6]. The cerebellum is relatively smaller than the cerebrum. It is responsible for motor control, i.e., systematic regulation of voluntary movements in living organisms with a nervous system. Due to variable size and stroke territory, ALI, lesionGnb, and LINDA methods fail to detect the small lesion region. Cerebellum is well-structured and welldeveloped in human beings as compared to other species [7]. The cerebellum has three lobes; an anterior, a posterior, and a flocculonodular. A round-shaped structure named vermis connects the anterior and posterior lobes. The cerebellum consists of an inner area of white matter (WM) and an outer greyish cortex, which is a bit thinner than that of the cerebrum. The anterior and posterior lobes assist in the coordination of complex motor movements. The flocculonodular lobe maintains the body's balance [4,8]. The brain stem, as the name states, is a 7-10 cm-long stem-like structure. It contains cranial and peripheral nerve bundles and assists in eye movements and regulations, balance and maintenance, and some essential activities such as breathing. The nerve tracks originating from the cerebrum's thalamus pass through the brain stem to reach the spinal cord. From there, they spread throughout the body. The main parts of the brain stem are midbrain, pons, and medulla. The midbrain assists in functions such as motor, auditory, and visual processing, as well as eye movements. The pons assists in breathing, intra-brain communication, and sensations, and medulla oblongata helps in blood regulation, swallowing, sneezing, etc. [9].

Brain imaging modalities
Three major methods (PET, CT, DWI and MRI) for brain tumors are widely used to analyze the brain structure.

Positron emission tomography
Positron emission tomography (PET) uses a special type of radioactive tracers. Metabolic brain tumor features such as blood flow, glucose metabolism, lipid synthesis, oxy-gen consumption, and amino acid metabolism are analyzed through PET. It is still considered as one of the most powerful metabolic techniques and utilizes the best nuclear medicine named as fluorodeoxyglucose (FDG) [33]. FDG is a widely used PET tracer in brain images. Nevertheless, FDG-PET images have limitations, e.g., an inability to differentiate between necrosis radiation and a recurrent high-grade (HG) tumor [34]. Moreover, during a PET scan, radioactive tracers can cause harmful effects to the human body, causing a post-scan allergic reaction. Some patients are allergic to aspartame and iodine. In addition, PET tracers do not provide accurate localization of anatomical structure, because they have a relatively poor spatial resolution as compared to an MRI scan [35].

Computed tomography
Computed tomography (CT) images provide more in-depth information than images obtained from normal X-rays. The CT scan has received widespread recommendation and adoption since its inception. A study [36] determined that in the USA alone, the annual CT scan rate is 62 million, with 4 million for children. CT scans show soft tissues, blood vessels, and bones of different human body parts. It uses more radiation than normal X-rays. This radiation may increase the risk of cancers when multiple CT scans are performed. The associated risks of cancers have been quantified according to CT radiation doses [37,38]. MRI can even help in evaluating structures obscured in a CT scan, and provides high contrast among the soft tissues, providing a clearer anatomical structure [39].

Magnetic resonance imaging
An MRI scan is used to completely analyze different bodyparts, and it also helps to detect abnormalities in the brain at earlier stages than other imaging modalities [40]. Hence, complex brain structures make tumor segmentation a challenging task [41][42][43][44][45][46][47]. This review discusses preprocessing approaches, segmentation techniques [48,49], feature extraction and reduction methods, classification methods, and deep learning approaches. Finally, benchmark datasets and performance measures are presented.

Diffusion weighting imaging
MRI sequences are utilized to analyze the stroke lesions based on the several parameters such as age, location and extent regions [50]. In the context of treatment, a computerized method might be utilized for accurate diagnosis of the disease progression rate [51]. The neuroscientists of cognitive, who frequently conduct research in which cerebral impairments are linked to cognitive function They observed Fig. 1 Datasets for brain tumor detection that segmentation of the stroke lesions is a vital task to analyze the total infected region of brain that provide help in the treatment process [52]. However, segmentation of the stroke lesions is a difficult task, because stroke appearance is change as the passage of time. The MRI sequence such as diffusion weighted imaging (DWI) and FLAIR are utilized for stroke lesions detection. In acute stoke stage DWI sequence highlight the infection part as a hyperintensity. The underperfusion region represents the mapping magnitude of the perfusion [53]. The dis-similarity among two regions might be considered as penumbra tissue. Stroke lesions appear in distinct locations and shapes. Different types of lesions are appeared in a variable size and shape and these lesions are not aligned with vascular patterns and more than one lesions might appeared on similar time. The size of the stroke lesions is in radii of the few millimeters and appears in a full hemisphere. The structure of the hemisphere is dissimilar, and its intensity might significantly vary within the infected region. Furthermore, automated stroke segmentation is difficult due to the similar appearance of the pathology such as white matter hyperintensities and chronic stroke lesions [54].

Evaluation and validation
In the existing literature, experimental results are evaluated on publicly available datasets to verify the robustness of algorithms.

Publicly available datasets
Several datasets are publicly available that are used by the researchers to evaluate the proposed methods. Some important and challenging datasets are discussed in this section. BRATS are the most challenging MRI datasets [55][56][57]. BRATS Challenge is published in different years with more challenges having 1 mm 3 voxels resolution. The detail of datasets is given in Fig. 1 as well as in Table 1.

Performance metrics
The performance measures play a significant role to compute the method's effectiveness. A list of performance metrics is provided in Fig. 2 [102] are applied for intensity normalization. Wiener filter is used to enhance the local and spatial information in medical imaging [103]. The widely utilized preprocessing methods are N4ITK [104] for the correction of bias field, median filter [104] for image smoothing, anisotropic diffusion filter [105], image registration [106], sharpening [107], and skull stripping through brain extraction tool (BET) [108].

Conventional methods
The conventional methods [46] are further categorized into the following: • Thresholding methods.
• Region growing methods.

Segmentation
Segmentation extracts the required region from input images. Thus, segmenting accurate lesion regions is a more crucial task [109]. As manual segmentation process is erroneous [110]; therefore, semi-and fully automated methods are utilized [46]. Segmentation of tumor region using semiautomated methods achieves acceptable outcomes over manual segmentation [111,112]. Semi-automated methods are further divided into three forms: initialization, evaluation, and feedback response [113,114].

Thresholding methods
The thresholding method is a basic and powerful method to segment the required objects [18] and the selection of an optimized threshold is a difficult task in low-contrast images. Histogram analysis is used to select threshold values based on image intensity [115]. Thresholding methods are classified into local and global. If high homogeneous contrast or intensity exists among the objects and background, then the global thresholding method is the best option for segmentation. The optimal threshold value can be determined by Gaussian distribution method [116]. These methods are utilized when the threshold value cannot be measured from the whole image histogram or single value of the threshold does not provide good results of segmentation [117]. In most cases, the thresholding method is applied at the first stage for segmentation and many distinct regions are segmented within the gray-level images as shown in Fig. 5.

Region growing (RG) methods
In RG approaches, image pixels form disjoint areas are analyzed through neighboring pixels, which are merged with homogeneousness characteristics based on pre-defined similitude criteria. The region growing might fail to provide better accuracy due to the partial volume effect [118,119]. To overcome this effect, MRGM is preferred [86,120]. The region growing with BA methods is also introduced [87].

Watershed methods
As  Table 3 lists the summary of feature extraction methods.

Deep learning methods
The variety of state of the art deep learning methodologies are used to learn the data in the medical domain [214] including CNN [215,216], Deep CNN, cascaded CNN [217], 3D-CNN [218], convolutional encoder network, LSTM, CRF [218], U-Net CNN [219], dual-force CNN [220] and WRN-PPNet [221]. The brain tumor classification problem has been solved by employing a LSTM model. In this method, input MRI images smooth using N4ITK and 5 × 5 Gaussian filter and passed as input to the four LSTM model. The LSTM model is constructed on the four hidden Units such as 200, 225, 200, 225, respectively. The performance of this model has been tested on BRATS (2012-2015 and 2018) series and SISS-2015  [189]. The 09-layer CNN model is trained from the scratch for classification of different types of tumors such as pituitary, glioma and meningioma. The method achieved an accuracy of the classification is 98.71% [223]. This model is trained from the scratch on publicly 696 weighted-T1 sequences. The model provides an accuracy of greater than 99% for tumor classification [224]. The existing methods are summarized in Table 5.
Although much work is done on deep learning methods, still there exist many challenges. The present methods do not  -achieve maximum results in the sub-structure of the tumor region. For example, if the accuracy of the complete tumor is increased, then the accuracy of the core and the enhanced tumor is decreased (as shown in Table 5).

Brain tumor detection using transfer learning
The manual detection of brain tumors is difficult due to asymmetrical lesions shape, location flexibility, and unclear boundaries. Therefore, a transfer-learning model has been suggested based on the super-pixel. The VGG-19 is a pretrained model that has been utilized for the classification of the different grades of the glioma such as high/low glioma. The method achieved 0.99 AUC on the brats 2019 series [232]. The three different types of pre-trained models i.e., VGG network, Google network and Alex network are employed on the brain datasets for the classification of glioma, pituitary and meningioma. In this method, augmentation methods are also employed on MRI slices to generalize the outcomes and reduced the overfitting prob-lem by increasing the quantity of the input data. After the experimental analysis using different pre-trained models, we conclude that VGG-16 provides greater than 98% classification accuracy [233]. The classification of brain tumors has been done using two different types of networks, i.e., visual attention network and CNN are utilized for classification of different types of brain tumor i.e., glioma, pituitary I and meningioma [234]. A pre-trained model i.e., VGG-16, Alex and Google net are investigated for the analysis of brain tumors. The frequency domain techniques have been applied on input slices to improve the image contrast. The contrast improved images are passed in the next phase. Where pretrained VGG-16 provides maximum classification outcomes [235]. The Laplacian filter with a multi-layered dictionary model is utilized for the recognition of brain tumors. The model performed better as compared to existing works [236]. The method consists of the three major steps such as pre-processing, augmentation of data, and segmentation and classification using transfer learning models. In which ResNet-50, DenseNet-201, MobileNet-v2 and Inceptionv3 are utilized to classify the brain lesions with 0.95 IoU [237]. The deep features are extracted from the transfer learning AlexNet model. The model has eight layers, five of which are convolutional and three of which are fully linked. The SoftMax layer has been employed for classification between the different types of brain lesions [238]. The transfer learning ResNet-50 model with average global pooling is utilized to reduce the gradient vanishing and overfitting issues. The performance of this model has been evaluated on three distinct types of brain imaging benchmark samples that contain 3064 input images. The method achieved an accuracy of the 97.08% that is maximum as compared to latest existing works [239]. A deep CNN was used in this study that based on transfer learning such as ResNet, Xception and Mobilenetv2 are utilized for the extraction of deep features has been for tumors classification using MRI images. This method achieved an accuracy of up to 98% [240]. In this method, Grab Cut method has been employed for segmentation of the brain lesions. Later hand-crafted such as LBP features dimension of 1 × 20 and HOG features dimension of 1 × 100 are extracted and serially fused to the deep features dimension of 1 × 1000 that are extracted from the pre-trained VGG-19 model and final fused features vector length of 1 × that is supplied to the different kind of classifiers. The experimental analysis proves that fused features vector provide good results as compared to existing work in this domain [16,187]. The global thresholding method is applied to segment the actual lesion region. After segmentation, texture features such as LBP and GWF are extracted from the segmented images. After that, the retrieved features are fused to form a single fused feature vector, which is then provided to the classifiers for differentiation between healthy and unhealthy images [26]. There are two key stages to the procedure. The brain lesions are enhanced and segmented using spatial domain approaches in the first stage, then deep information's are extracted using pre-trained models, i.e., Alex and Google-network and score vector is achieved from softmax layer that is supplied to the classifiers such as for discrimination between the glioma/non-glioma images of brain. The Brats series dataset was used to test this technique's efficiency [241]. For brain tumor segmentation, the superpixel approach has been suggested. From the segmented images, Gabor wavelet information are retrieved and given to SVM and CRF for discrimination between the healthy/unhealthy MRI images [242].The transfer learning models such as inceptionv3, densenet-201, and to form a single vector, extracted features are merged serially and passed to softmax for tumor classification. Furthermore, different dense blocks of the densenet201 are extracted and classify the brain tumor using softmax. The approach had a 99% accuracy rate. The evaluation outcomes clearly state that the fused vector outperformed as compared to the single vector [243]. A novel U-net model with the RESnet model has been trained on the input MRI images. The classifiers are fed the salient features derived from its pictures. This method has been tested on BRATS 2017, 2018 and 2019 datasets [244]. The tumor region is localized on Flair sequences of brats 2012 series. The skull is removed from of the input pictures, and a noise-reduction filter is applied bilaterally. During the segmentation, texton features are recovered from the input images using the superpixel approach. For brain tumor classification, the leave out validation technique is used. This strategy yielded an 88 percent dice score [245].
The deep segmentation has been designed that contains two major parts such as encoder and decoder. The spatial information is extracted using a CNN in the encoder section. For determining the whole probability map resolution, the semantic mappings information is entered into the decoder component. On the basis of U-network distinct CNN networks such as ResNetwork, dense network and Nas-network are utilized for features extraction. This model has been tested successfully on Brats-2019 series. The method achieved dice scores of 0.84 [246]. The wavelet homomorphic filter has been employed for noise removal. The tumor infected region has been localized using improved YOLOv2 model [230]. The summary of the transfer learning methods is mentioned in Table 6.

Brain tumor detection using quantum machine learning
Superposition of quantum states/parallelism/entanglement can all be used to establish quantum computer supremacy [258]. However, exploring entanglement of quantum features for efficient computation is a difficult undertaking due to a shortage of computational resources for execution of quantum algorithms. With the progress of quantum techniques, classical computers based on quantum theory and influenced through qubits are no longer able to fully exploit the benefits of quantum state and entanglement. QANN has been found to be effective in a variety of computer tasks, including classification and pattern recognition due to the intrinsic properties supplied by quantum physics [259]. On the other hand, quantum models based on genuine quantum computers use big bits of the quantum/qubits as a simple representation of matrix and the linear functions. However, the computational complexity of the quantum-inspired neural network (QINN) designs increases several fold due to complicated and timeconsuming back-propagation quantum model [260]. The automatic segmentation of brain lesions from I (MRI), which removes the onerous manual work of human specialists or radiologists, greatly aids brain tumor detection. Manually, brain tumor diagnosis, on the other hand, suffers from large variances in size, shape, orientation, illumination variations, greyish overlaying, and cross-heterogeneity. Scientists in the computer vision field have paid a lot of emphasis in recent years to building robust and efficient automated segmentation approaches. The current research focuses on a unique quantum fully supervised learning process which is defined by qutrits for timely and effective lesions segmentation. The proposed work's main goal is to speed up the QFSconvergence Net's and make it appropriate for computerized segmentation of the brain lesions without the need for any learning/supervision. To leverage the properties of quantum correlation, suggested a quantum fully self-supervised neural network (QFS-Net) model uses qutrits/three states of quantum for segmentation of the brain lesions [261]. The QFS-Net uses a revolutionary fully supervised qutrit-based counter propagation method to replace the sophisticated quantum back-propagation method that utilized in supervised QINN networks. This approach allows for iterative quantum state that propagates among the layers of network.

Limitations of existing's machine/deep learning methods
In this survey, recent literature regarding the detection of brain tumors is reviewed, and it is indicated that there is still room for improvement. During image acquisition, noise is included in MRI, and noise removal is an intricate task [2, [262][263][264]. Accurate segmentation is a difficult task [265], as brain tumors have tentacles and diffused structures [43,193,220,266]. Selecting and extracting optimal features and appropriate number of training/testing samples for better classification is also an important task [191,192]. Deep learning models are gaining attention as the learning of features is accomplished automatically; however, they require high computing power and large memory. Therefore, still there is a need to design a lightweight model that provides high ACC in less computational time. Some existing machine learning methods with their limitations are mentioned in Table 7.
The following are the main challenges of brain tumor detection.
The glioma and stroke tumors are not well contrasted. It consists of tentacle and diffused structures that make segmentation and classification processes more challenging [270].
A small volume of tumor detection is still a challenge as it can be detected as a normal region [269,273].
Some of the existing methods work well for only a complete tumor region and do not provide good results for other regions (enhanced, non-enhanced) and vice versa [267,271,274].

Research findings and discussion
After a comprehensive review of the state-of-the-art exiting methods, the following challenges are found: • The size of a brain tumor grows rapidly. Therefore, tumor diagnosis at an initial stage is an exigent task. • Brain tumor segmentation is difficult owing to the following factors. • MRI image owing to magnetic field fluctuations in the coil.
• Gliomas are infiltrative, owing to fuzzy borders. Thus, they become more difficult to segment [43]. • Stroke lesion segmentation is a very intricate task, as stroke lesions appear in complex shapes and with ambiguous boundaries and intensity variations. • The optimized and best feature extraction and selection is another difficult process inaccurate classification of brain tumors.

Conclusion
The accurate brain tumor detection is still very demanding because of tumor appearance, variable size, shape, and structure. Although tumor segmentation methods have shown high potential in analyzing and detecting the tumor in MR images, still many improvements are required to accurately segment and classify the tumor region. Existing work has limitations and challenges for identifying substructures of tumor region and classification of healthy and unhealthy images. In short, this survey covers all important aspects and latest work done so far with their limitations and challenges. It will be helpful for the researchers to develop an understanding of doing new research in a short time and correct direction.
The deep learning methods have contributed significantly but still require a generic technique. These methods provided better results when training and testing are performed on similar acquisition characteristics (intensity range and resolution); however, a slight variation in the training and testing images directly affects the robustness of the methods. In future work, research can be conducted to detect brain tumors more accurately, using real patient data from any medium (different image acquisition (scanners). Handcrafted and deep features can be fused to improve the classification results. Similarly, lightweight methods such as quantum machine learning play significant role to improve the accuracy and efficacy that save the time of radiologists and increase the survival rate of patients.

Declarations
Conflict of interest There is no grant received from any resources. All authors declare that they have no conflict of interest.

Research involving human participants and/or animals It is declared that research has not involved any human participants and animals.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecomm ons.org/licenses/by/4.0/.