A Survey on Various Machine Learning Techniques for an Efficient Brain Tumor Detection from MRI Images

A Survey on Various Machine Learning Techniques for ABSTRACT - On account of the uncontrolled and quick growth of cells, Brain Tumor (BT) occurs. It may bring about death if not treated at an early phase. Brain Tumor Detection (BTD) has turned out to be a propitious research field in the current decennia. Precise segmentation along with classification sustains to be a difficult task in spite of several important efforts and propitious results in this field. The main complexity of BTD emerges from the change in tumor location, shape, along with size. Providing detailed literature on BTD via Magnetic Resonance Imaging (MRI) utilizing Machine Learning (ML) methods to aid the researchers is the goal of this review. Diverse datasets are mentioned which are utilized most often in the surveyed articles as a prime source of Brain Disease (BD) data. Furthermore, a concise epitome of diverse segmentation methods that are utilized in diagnosing BDs has been offered. Lastly, an outline of key outcomes from the surveyed articles is exhibited, and several main problems related to ML-centred BD diagnostic methodologies are elucidated. The most precise method to detect diverse BDs can be engaged for future advancement via this study.


░ 1. INTRODUCTION
An intricate human body organ that operates via billions of cells is called a Brain. Normal brain activities are impacted by these cells and as well damage normal cells. Therefore, one of the main grounds for death in adults across the globe is BT [19]. Via the early BT's detection, the lives of millions can be saved [41]. Initial BTD can guide the patients to get on-time treatment and aid to augment the patient's life expectancy [26]. Typically, by means of brain imaging methods like Positron Emission Tomography (PET), Single-Photon Emission Computed Tomography (SPECT), Computed Tomography (CT), MRI, and Magnetic Resonance Spectroscopy (MRS), the BT's early diagnosis is carried out that is utilized to offer information regarding the size, location, shape, and kind of BT to help out in the diagnosis [4]. Since MRI scans give much information about the images within the human tissues in views of '3' dimensions, MRI is the most noteworthy one of all [34]. To generate images of human tissues, MRI wields radiofrequency signals with a powerful magnetic field. So, the tumor is revealed more evidently in MRI, which aids in the process of further treatment [42]. However, it is extremely tough to identify the tumor's existence owing to the intricate brain structure that differs with age along with pathological history. During the recent decades, noteworthy research in the area of BT diagnosis has been performed by several researchers. But, there are limited applications of the existent works. Even though a wide number of works have been implemented, clinicians still rely on the tumor's manual projection, perhaps on account of the deficiency of a link betwixt clinicians and researchers [32]. Therefore, a survey of the largely significant prevailing BT diagnosis techniques is displayed in this work. MRI BT diagnosis with conventional ML methods is the A Survey on Various Machine Learning Techniques for primary focus of this survey. Even though numerous reviews are obtainable in the literature that particularly concentrates on one specific process like segmentation, classification, or diagnosis, but, this paper examines a detailed summary of the complete BT diagnosis system regarding tumor detection, segmentation, along with classification. Also, the advantages along with disadvantages of the conventional ML methodologies are enveloped in this study.
The remaining of this work is arranged as: A comprehensive review of diverse algorithms to execute BTD is provided in section 2. The outcomes along with comparison of the diverse methods for BTD are exhibited in section 3. A few of the difficulties along with opportunities in this area together with recognition of possible future guidelines are deduced in section 4.

░ 2. LITERATURE REVIEW
Several studies have emerged in the field of BTD in the past few decennia. A concise review of diverse algorithms which has been performed in this field is offered in this section. The diverse methods for BTD are surveyed in section 2.1, manifold segmentation algorithms for BTD are elucidated in section 2.2, and MRI image-centered BTD utilizing ML is explicated in section 2.3.

Survey of different Techniques for Brain Tumor Detection
On account of the intricacy of scrutinizing along with diagnosing it as of the MR images, BTD has been one of the primary concerns. Therefore, numerous algorithms are utilized for effective BTD. A survey of different BT methodologies proposed by authors for efficient BTD is displayed in Table 1.

░ Table 1. Survey of various BTD models
Neelum Noreen et al. [25] put forward a technique of features extraction together with concatenation for the earlier diagnosis of BTs. Primarily, as of the pre-trained Inception-v3 model, the features from varied Inception modules were extracted together with concatenated these features for BT classification. Next, these features proceeded to the Softmax Classifier (SC) that classified the BT. Secondly, to extract the features from a variety of DenseNet blocks, pre-trained DensNet201 was utilized. Next, these features were concatenated along with inputted into the SC for the BT classification. With Inception-v3 along with DensNet201, the technique generated testing accuracy of 99.34 % and 99.51%, respectively. It also accomplished the greatest performance in the BTD.
Amran Hossain et al. [2] suggested the BT's detection via the YOLOv3 Deep Neural Network (DNN) in a portable Electromagnetic (EM) imaging scheme. The technique's detection performance was examined with the diverse image datasets. The detection accuracy of 95.62% and F1 score of 94.50% were attained by the methodology. 96.74% of training accuracy and 9.21% of validation losses were acquired. Via YOLOv3, the tumor detection with its site in diverse cases as of the testing images was scrutinized, which evinced its possible EM head imaging system. Yet, the scheme had skip connections problems.

Segmentation Techniques for MRI Brain Tumor Detection
As of the images, segmentation extracts the necessary region. So, it is a decisive task to segment exact lesion regions as of the brain MRI images. Semi-and fully automated techniques are wielded since the manual segmentation process is inaccurate. When analogized to manual segmentation, the tumor region's segmentation utilizing automated techniques attains satisfactory results. Hence, a diverse segmentation method's survey recommended by a variety of authors for effective BTD is shown in Table 2

MRI image-based Brain Tumor Detection Using Machine Learning
A process of training a computer to employ its earlier experience to resolve an issue offered to it is termed ML.
Owing to the present accessibility of inexpensive computing power along with memory, the notion of ML's application in diverse fields for solving issues quicker than humans has received considerable interest. Conversely, Deep Learning is a sub-field of ML. Usually, the ML algorithms are divided into '2' categories Supervised Learning (SL) and Unsupervised Learning (UL). Here, an outline of the most effective and familiar methods of ML with their outcomes is displayed.

Supervised Machine Learning Techniques for Brain Tumor Detection
SVM, Adaptive-Network-centered Fuzzy Inference System (ANFIS), KNN, Naïve Bayes (NB), CNN, DNN, etc., are the classification methods in SL. The ML-centered SL methods are surveyed in this part and listed in Table 3 and Table 4.

Unsupervised Machine Learning Techniques for Brain Tumor Detection
The clustering methods and association methods come under UL. The ML-centered UL methodologies are inspected in this part. Md Khairul Islam et al. [6] applied an enriched BTD scheme grounded on the Template-centered K-means (TK) algorithm with super pixels together with PCA that effectively detected the human BTs in lesser Execution Time (ET). Initially, utilizing both super pixels and PCA, necessary features were extracted. Next, utilizing a filter, image enrichment was executed. Lastly, via the TK-means clustering algorithm, image segmentation was performed. The experimental outcomes exhibited that the system accomplished superior accuracy along with a decreased ET than other prevailing systems meant for the BT's detection in an MR image. However, the system had higher time complexity.
Khurram Ejaz et al. [35] initiated UL with a feature approach for BTS utilizing MRI. Here, the image's utmost and least intensities had been attuned, which emphasized the tumor portion; then, a thresholding function was executed that localized the tumor region. Next, unsupervised clusters like Kmean were employed for the tumor's separation from boundaries. However, the scheme had the incapability to extract boundary features for similar regions.
Xinheng Wu et al. [10] investigated an unsupervised BTS technique called Symmetric Driven Generative Adversarial Network (SD-GAN). SD-GAN scheme was trained and learned a non-linear mapping betwixt the left and right brain images, along with variability of the brains was forecasted.

░ 3. RESULTS AND DISCUSSION
Here, utilizing diverse ML methods, the BTD's performance along with the efficient comparison regarding the accuracy, sensitivity, specificity, and SA, along with their comparative investigation was done centered on a variety of datasets. the method EKF-SVM attains 97.04%, respectively, regarding sensitivity and specificity. Nearly all algorithms acquired better performance rates that overall vary from 92%-100%. So, it is deduced that most of the ML algorithms work well for BTD.

░ 4. CONCLUSION
In the current decennia, there has been increased attention towards BTD. In the past few years, the expansion of several novel BTD algorithms has been observed. Utilizing ML methodologies, this work offers a detailed review of current advancements in BTD. To attain the objectives of this survey, the benefits and drawbacks are discussed comprehensively. Centered on the database and recognition accuracy, the algorithm's outcomes are enlisted. Regarding their accuracy and complexity, it is helpful for researchers to create algorithms. However, during training, most of the ML models are time-consuming. It is indispensable to train the model in a short time and ameliorate real-time ability. Besides the probability of inhomogeneity of tumorous tissue, the diversity of the shape and intensity of tumors are the most vital restrictions that make BTS a difficult task. Hence, it is crucial to construct a pre-eminent model to segment the tumor as of MRI images.

░ 5. ACKNOWLEDGMENTS
Thanks to the supervisor who have supported towards development of the survey.