![Different types of brain tumors [38]. 138](https://www.researchgate.net/profile/Ramin-Ranjbarzadeh/publication/366110391/figure/tbl1/AS:11431281112931124@1673625718125/Different-types-of-brain-tumors-38-138_Q320.jpg)
Content uploaded by Ramin Ranjbarzadeh
Author content
All content in this area was uploaded by Ramin Ranjbarzadeh on Jan 13, 2023
Content may be subject to copyright.
Brain Tumor Segmentation of MRI Images: A Comprehensive 1
Review on the Application of Artificial Intelligence Tools 2
Ramin Ranjbarzadeh1, Annalina Caputo2, Saeid Jafarzadeh Ghoushchi3, Erfan Babaee Tirkolaee4,*, Malika 3
Bendechache5 4
1 School of Computing, Faculty of Engineering and Computing, Dublin City University, Ireland. 5
ramin.ranjbarzadehkondrood2@mail.dcu.ie (corresponding author) 6
2 School of Computing, Faculty of Engineering and Computing, Dublin City University, Ireland. 7
annalina.caputo@dcu.ie 8
3 Faculty of Industrial Engineering, Urmia University of Technology, Urmia, Iran. 9
s.jafarzadeh@uut.ac.ir 10
4 Department of Industrial Engineering, Istinye University, Istanbul, Turkey. 11
erfan.babaee@istinye.edu.tr (corresponding author) 12
5 Lero & ADAPT Research Centres, School of Computer Science, University of Galway, Ireland . 13
malika.bendechache@universityofgalway.ie 14
15
Abstract 16
Background 17
Brain cancer is a destructive and life-threatening disease that imposes immense negative effects on patients’ 18
lives. Therefore, detection of brain tumors at an early stage improves the impact of treatments and increases 19
patients' survival rate. However, detecting brain tumors in their initial stages is a demanding task and an 20
unmet need. 21
Methods 22
The present study presents a comprehensive review of the recent Artificial Intelligence (AI) methods of 23
diagnosing brain tumors using MRI images. These AI techniques can be divided into Supervised, 24
Unsupervised, and Deep learning (DL) methods. 25
Results 26
Diagnosing and segmenting brain tumors usually begin with Magnetic Resonance Imaging (MRI) on the 27
brain since MRI is a noninvasive imaging technique. Another existing challenge is that the growth of 28
technology is faster than the rate of increase in the number of medical staff who can employ these 29
technologies. It has resulted in an increased risk of diagnostic misinterpretation. Therefore, developing 30
robust automated brain tumor detection techniques has been studied widely over the past years. 31
Conclusion 32
The current review provides an analysis of the performance of modern methods in this area. Moreover, 33
various image segmentation methods in addition to the recent efforts of researchers are summarized. 34
Finally, the paper discusses open questions and suggests directions for future research. 35
Keywords: Brain Tumor, Artificial Intelligence, Tumor Segmentation, Tumor Classification, MRI 36
Modalities. 37
38
1. Introduction 39
Brain is an important organ containing one hundred billion nerve cells or neurons. According 40
to the reports, brain tumors are the 10th main cause of mortality among the adults and children for 41
both genders in developed countries [1], [2]. It is anticipated that the incidence of primary brain 42
tumors will cause 18280 deaths in adults in the USA in 2022 [3]. Brain tumors, known as 43
intracranial tumors, include a diverse set of cancerous cells that start in the intracranial tissues of 44
the brain and can range in malignancy from benign to advanced [4],[5]. Brain tumor begins in case 45
of cell division rates increase and multiply uncontrollably. Any portion of the brain or skull can 46
develop a brain tumor, including the brain's protective lining, skull base, brainstem sinuses, nasal 47
cavity, and many other places [6]. There are more than 150 kinds of brain tumors. The two basic 48
classifications of brain tumors are cancerous and noncancerous [7], [8]. 49
The brain is made up of several cell types, each with its own special characteristics. It is 50
impossible to generalize results from malignancies in other organs to those arising in the brain [9]. 51
The unique biology and microenvironment of the brain is the main aspect of brain cancers. Each 52
form of tumor has its own biology, course of therapy, prognosis outlook, and a different set of risk 53
factors [10], [11], and this makes the brain tumor classification difficult to describe. Pressure in 54
the head caused by a brain or spinal cord tumor is a common sign of brain cancer. People with 55
brain tumors are more likely to have specific symptoms, such as exhaustion, nausea, or discomfort. 56
The other side effects of living with brain tumors and brain cancer are fever, rash, and increased 57
pulse. The experts can link signs and symptoms to describe the problem with more certainty. 58
However, brain tumors do not always cause symptoms [12], [13]. 59
Diagnosing a brain tumor involves three different tests and procedures, including imaging tests, 60
neurological exams, and biopsy. The most common and well-established method to diagnose brain 61
cancers is using Magnetic Resonance Imaging (MRI) [14]. During an MRI scan, a dye can be 62
injected into a vein. Experts assess the tumor and make treatment plans based on the MRI scan 63
elements, e.g., perfusion MRI, functional MRI, and magnetic resonance spectroscopy. In some 64
circumstances, further imaging tests like Positron Emission Tomography (PET) and Computed 65
Tomography (CT) are used in combination with MRI. Problems in any area can signify which part 66
of the brain is affected by a tumor [15], [16]. In this case, neurological tests can help the expert for 67
a better diagnosis. During the neurological examination, the expert checks the hearing, vision, 68
coordination, balance, strength, and reflexes of the patient. For a more precise diagnosis, a biopsy 69
is utilized. In this procedure, a sample of abnormal tissue collects and examined under the 70
microscope [17], [18]. 71
If the practitioners can diagnose the disease early, it can result in a timely treatment which 72
increases the likelihood of survival. However, because tumor regions frequently have unclear 73
morphological structures, the identification of malignancies can be challenging [19]. Physicians 74
have recently used Computer-Aided Diagnostic (CAD) tools to aid in the diagnosis of cancer more 75
accurately [20]–[23]. These innovative methods of brain imaging have increased the detection ratio 76
of brain tumors [24]. Due to the requirement for intelligence in CAD systems, significant changes 77
have occurred recently, and Artificial Intelligence (AI) is integrated with CAD to reduce the 78
recognition time and system memory requirements [25], [26]. Also, it aids in the development of 79
a useful knowledge-based design system. Comparing the new AI assistance with the traditional 80
CAD system, one can realize that the new AI assistance, when combined with CAD, proves to be 81
more efficient. The accuracy and algorithms of AI systems have improved as a result of the 82
growing application of Deep Learning (DL) in medical research and the growth of big data 83
analytics. For example, a radiologist can apply the latest AI technology and advanced computer-84
assisted detection and diagnosis to collect more details about the normal and tumor tissues [27]. 85
Furthermore, it is possible to assess the patient's status by employing CADs and analyzing imaging 86
and/or non-imaging patient data. Over the last decades, this technology significantly influenced 87
early cancer detection and its timely treatment [28]. 88
This review has summarized more than 100 scientific research papers from 2015- 2022 (until 89
1st September 2022). To find out the number of investigations on brain tumor diagnosis through 90
supervised learning, unsupervised learning, and DL models a statistical report is provided based 91
on the “Scopus” database. The keywords searched in this database were “brain tumor” AND “name 92
of the technique (e.g., Random Forest)” OR “brain cancer” AND “name of the technique”. 93
This paper focuses on reviewing the studies that applied AI techniques for brain tumor 94
segmentation. Abbreviations used in this paper are referenced in Table 1. 95
The rest of the paper is organized as follows: Detailed images of body organs in Magnetic 96
Resonance Imaging (MRI) is described in section 2. Different types of tumors and their 97
characteristics are implied in section 3. In section 4, more details about supervised and 98
unsupervised techniques are represented. Next, some Deep learning models applied in the field of 99
brain tumor segmentation are discussed in section 5. In the next step, some top databases for the 100
brain tumor segmentation are represented in section 6. Then, performance measures are described 101
in section 7. Finally, discussion and conclusion parts are provided in sections 8 and 9. 102
103
104
Table 1. List of abbreviations. 105
Description Abbreviation Description Abbreviation
Magnetic Resonance Imaging MRI Cerebrospinal Fluid's CSF
Artificial Intelligence AI Glioblastoma Multiforme GBM
Positron Emission Tomography PET Low-Grade Glioma LGG
Computed Tomography CT High-Grade Glioma HGG
Computer-Aided Diagnostic CAD Radio Frequency RF
Repetition Time TR Time To Echo TE
Fluid Attenuated Inversion
Recovery FLAIR Support Vector Machine SVM
Artificial Neural Network ANN Random Forest RF
K-Nearest Neighbors Algorithm KNN Linear Discriminant Analysis LDA
Genetic Algorithm GA Maximum Marginal Hyperplane MMH
Social Ski Driver SSD Kernel Support Vector Machine KSVM
Random Decision Forest RDF Gaussian Mixture Model GMM
Decision Tree DT Random Forest Classifier RFC
Proton density PD Chemical shift imaging CSI
Whale Optimization Algorithm WOA An Adaptive Artificial Neural
Network AANN
Fuzzy-C-Mean FCM Naive Bayes Classifier NBC
Harmony-Crow Search HCS Particle Swarm Optimization PSO
Learning Vector Quantization LVQ Self-Organizing Maps SOM
Principal Component Analysis PCA Adaptive Kernel Fuzzy C-Means AKFCM
Contrast Enhanced Fuzzy C-Means CEFCM Pixel-Based Voxel Mapping
Technique PBVMT
Multiscale Fuzzy C-Means MsFCM Gray-Level Co-Occurrence Matrix GLCM
Edge Adaptive Total Variation
Denoising Technique EATVD Imaging Mass Spectrometry IMS
Hierarchical Cluster Analysis HCA Region Of Interest ROI
Association Allotment Hierarchical
Clustering AAHC Density-Based Spatial Clustering of
Applications with Noise DBSCAN
Gustafson-Kessel (G-K) Convolutional Neural Network CNN
Enhancing Tumor ET Whole Tumor WT
Tumor Core TC Recurrent Neural Network RNN
Long Short-Term Memory LSTM Generative Adversarial Networks GAN
Residual Cyclic Unpaired Encoder-
Decoder Network RescueNet Sailfish Political Optimizer SPO
Deep Belief Networks DBN Stacked Sparse Autoencoder SSAE
Reinforcement Learning RL Gated Recurrent Unit GRU
Density-Based Spatial Clustering DBSCAN Gaussian Mixture Models GMM
Earthworm Optimization Algorithm EWA Monarch Butterfly Optimization MBO
Harris Hawks Optimization HHO Moth Search Algorithm MSA
Hunger Games Search HGS Runge Kutta Optimizer RUN
Slime Mould Algorithm SMA Colony Predation Algorithm CPA
2. Magnetic Resonance Imaging In Brain Tumor Detection 106
Clinicians can plan the most effective and practical treatment for patients involved with brain 107
cancer by obtaining information from several restorative diagnostic imaging technologies, 108
including MRI, PET, and CT [29]. However, better images of organs and soft tissues can be 109
produced using MRI. Using radio waves and strong magnetic fields, MRI releases detailed images 110
of body organs [30]. Compared to a CT scan or X-rays, MRI provides clearer images and it is a 111
better option when doctors should see soft tissues. Brain tumor location and size are determined 112
using imaging methods like MRI [8]. MRI images typify significant data about tissue 113
characteristics, for example, Proton density (PD), spin-lattice (T1), and spin-spin (T2) relaxation 114
durations, chemical shift imaging (CSI), and flow velocity. These facts allow for a more accurate 115
representation of brain tissue [31], [32]. MRI scans can acquire images with various contrasts using 116
various procedures or acquisition parameters. T1 weighted images with contrast material 117
Gadolinium (T1c) aid in distinguishing tumor borders from surrounding normal tissues. T2 118
weighted (T2) images are typically employed to provide an underlying assessment, identify 119
different tumor types, and distinguish cancers from normal tissues. No enhanced tumors are seen 120
using the T2 weighted scan in axial viewing with FLAIR. Given these unique characteristics, MRI 121
provides a decision-making advantage in investigations of brain tumors [12]. Table 2 indicates the 122
four types of MRI images. 123
124
Table 2. MRI image modalities [34]. 125
Type Feature
T1
T1-weighted MRI
Calculate the tissue's T1 (longitudinal) relaxation time.
Brighter tissue has shorter relaxation times.
T2
T2-weighted MRI
Calculate the tissue's T2 (transverse) relaxation time.
Longer relaxation times result in brighter tissue
T1c
T1 weighted images with contrast material Gadolinium.
The signal for tumor increase
FLAIR
Fluid Attenuated Inversion Recovery MRI.
Cerebrospinal fluid's (CSF) bright signal is suppressed.
Can more accurately identify small hyper-intense lesions.
3. Brain Tumor Types 126
Brain tumor, sometimes referred to as an intracranial tumor, is an abnormal lump where cells 127
are amassed to reproduce uncontrollably [35]. Currently, over 120 types of brain tumors are 128
detected, two basic types of which are primary and metastatic. Primary brain tumors, also known 129
as Meningioma, are tumors that develop from the brain's tissues or its immediate surroundings and 130
account for more than 30% of all brain tumors [36]. Glial (consisting of glial cells) and non-glial 131
(formed on or in the brain structures; i.e., nerves, glands, and blood vessels) primary tumors are 132
classified as benign or malignant. Those tumors developed in other parts of the body, like the lungs 133
or breast, and spread to the brain typically through the blood flow are referred to as metastatic 134
brain tumors [37]. Malignant tumors with metastases are regarded as cancer. There are several 135
types of brain tumors. Table 3 represents the most common brain tumors. 136
137
Table 3. Different types of brain tumors [38]. 138
Brain Tumor Types Subtype
1
Gliomas
Astrocytoma
2 Pilocytic Astrocytoma (grade I)
3 Diffuse Astrocytoma (grade II)
4 Anaplastic Astrocytoma (grade III)
5 Glioblastoma Multiforme (grade IV)
6 Oligodendroglioma (grade II)
7 Anaplastic Oligodendroglioma (grade III)
8 Ependymoma (grade II)
9 Anaplastic Ependymoma (grade III)
10 Craniopharyngioma
11 Epidermoid
12 Lymphoma
13 Meningioma
14 Schwannoma (neuroma)
15 Pituitary adenoma
16 Pinealoma (Pineocytoma, Pineoblastoma)
139
One of the most common types of brain cancer is Gliomas and account for almost 33% of all 140
brain cancers [37], [39]. As gliomas frequently mix with healthy brain tissue and develop within 141
the substance of the brain, they are sometimes referred to as intra-axial brain tumors. Glioblastoma, 142
also named Glioblastoma Multiforme (GBM) and is challenging for experts to diagnose and cure. 143
Fig. 1 shows the Glioblastoma imaging. 144
145
146
Fig. 1. MRI image of Glioblastoma. 147
Glioblastoma often has a blend of cell grades and changes synchronous to their growth. The 148
features of tumors that appear under a microscope and the aggressiveness of the tumor make it 149
possible for the experts to recognize tumor types. For example, if the grades are low, it means they 150
are least aggressive, and if the grades are high, it indicates they are most aggressive. You can see 151
the characteristics of Glioma scales in Table 4. 152
153
Table 4. Glioma grades and their characteristics [37]. 154
Grade Characteristic
I
Near to normal appearance
Least malignant
Slow growing cells
Commonly indicates long
-
term survival
II
Cells growing fairly slowly
Fairly abnormal appearance
Able to attack nearby tissue
In some cases recur as a higher grade
III
Actively creating abnormal cells
Abnormal appearance
Infiltrate normal tissue
Tend to recur, often as a higher grade
IV
Rapidly reproducing abnormal cells
Highly abnormal appearance
Area of dead cells (necrosis) in center
Form new blood vessels to continue
growth
155
In glioma, grades I and II (Table 4) are categorized as Low-Grade Glioma (LGG), and grades 156
III and IV are High-Grade Glioma or HGG [40]. Taking the best procedure for treatment depends 157
on the early diagnosis of Glioma and its grade [8], [37], and MRI imaging is one of the bests 158
screening methods for diagnosing Glioma. In this screening method, the protons which are 159
randomly oriented inside the water nuclei of the tissue are brought into alignment using a strong, 160
uniform, external magnetic field. The subsequent disruption of this alignment (or magnetization) 161
is caused by the addition of an external Radio Frequency (RF) energy. Through a variety of 162
relaxation mechanisms, the nuclei return to their resting alignment and release RF energy in the 163
process. The emitted signals are evaluated after a certain amount of time has passed from the first 164
RF. The signal from each place in the imaged plane is converted using the Fourier transform into 165
the relevant intensity levels, which are then represented as shades of gray in a matrix of pixels 166
[41]. Different forms of images can be produced by altering the order in which RF pulses are 167
delivered and collected. The T1- and T2-weighted scans are very popular MRI sequences. Short 168
Time to Echo (TE) and Repetition Time (TR) is utilized to create T1-weighted images. T1 169
characteristics of tissue are primarily responsible for determining the contrast and brightness of 170
the image. On the other hand, longer TE and TR times are used to create T2-weighted images. The 171
T2 characteristics of the tissue determine the contrast and brightness in these images. The Fluid 172
Attenuated Inversion Recovery (Flair) sequence is a third frequently utilized sequence. The TE 173
and TR timings of the Flair sequence are significantly longer than those of a T2-weighted image 174
[42]. Fig. 2 indicates the MRI image T1, T2, FLAR modalities. 175
176
177
Fig. 2. Some examples of MRI images including T1, T1ce, T2, and Flair. 178
Table 5 shows the frequently-used MRI sequences, along with an estimate of their TR and TE 179
times for different tissues. 180
181
Table 5. Most common sequences of the MRI images [43]. 182
Tissue T1 T2 FLAIR
White matter Light Dark gray Dark gray
CSF Dark Bright Dark
Fat Bright Light Light
Cortex Gray Light gray Light gray
Inflammation Dark Bright Bright
4. Brain Tumor Segmentation 183
Segmentation means grouping regions or pixels of images into many coherent subregions and 184
categorizing each subregion into one of the specified classes based on the extracted features, such 185
as color or texture attributes [44]–[46]. Segmentation is a form of image compression and has 186
extensive applications in the development of CAD that works based on radiological images such 187
as MRI. Generally, image segmentation can be categorized into two main groups: supervised and 188
unsupervised [47]–[49]. Unsupervised segmentation approaches include thresholding, edge 189
detection, graph cutting, and deformation to define the boundaries of the target item in the image 190
[50], [51]. In contrast, supervised segmentation techniques apply training samples to include prior 191
knowledge about the image processing problem. Fig. 3 shows the categories of supervised and 192
unsupervised approaches. 193
194
Fig. 3. Brain tumor segmentation approaches. 195
To find out the number of investigations on brain tumor diagnosis through supervised and 196
unsupervised learning, a statistical report is provided based on the “Scopus” database. The 197
keywords searched in this database were “name of the method (e.g., SVM)” AND “brain tumor” 198
OR “name of method” AND “brain cancer.” The date of publication was set from 2015- 2022 199
(until 1st September 2022). Fig. 4 shows the number of publications that used Machine Learning 200
(ML) approaches to diagnose brain cancer. According to Fig. 4, “Support Vector Machine 201
(SVM),” with 1136 papers published between 2015 to 2022, is the most used approach for brain 202
tumor classification or segmentation. This approach is employed in two ways, directly applied to 203
databases or used in a hybrid with other techniques. SVM, more than 90% of the time, indicated a 204
high accuracy rate, and this is the main reason for the highest number of publications based on this 205
approach. The second approach that is applied to diagnose brain tumors is “Artificial Neural 206
Network (ANN)”. The number of investigations conducted based on this approach is about 846 207
papers. Random forest is the third popular approach with 834 papers. In the following, the ML 208
approaches are introduced briefly and the several papers that applied the approach are discussed. 209
210
Fig. 4. Number of publications between 2015 to 2022 for brain tumor segmentation using supervised and 211
unsupervised learning techniques. 212
213
4.1 Supervised Learning 214
Supervised learning is a popular branch of ML algorithms and commonly is referred to as 215
supervised ML. In this approach, labeled datasets are utilized to train computers in order to 216
properly categorize data or predict outcomes [52], [53]. The model modifies its weights until the 217
model is properly fitted, which is a part of the cross-validation process. A training set is used in 218
supervised learning to instruct patterns to produce optimal results. This training dataset has 219
accurate input data and outputs (labels), enabling the model to develop through time. The 220
algorithm evaluates precision using the loss function and modifies it to minimize the error. 221
Supervised learning techniques assist in finding large-scale solutions to real-world issues, 222
including in the medical field. Various methods of supervised learning have been applied to 223
diagnose a brain tumor. In the following, a list of techniques is provided. 224
The most important benefit of these methods is that they allow to produce a data output or 225
collect data from the prior experience. The main disadvantage of these models is inability to 226
correctly classify an input data that was not belonged to any classes in the training data. 227
4.1.1 K-nearest Neighbors Algorithm 228
The K-Nearest Neighbors (KNN) algorithm is a supervised learning classifier that employs 229
proximity for classifications or predictions about the grouping of a data point [54]. Although it can 230
be applied to classification or regression issues, it is commonly employed as a classification 231
method since it relies on the concept that similar points can be discovered near one another [55]. 232
It is also known as a lazy learner algorithm. KNNs have been common in brain cancer 233
segmentation, and the results of studies had different accuracy rates. For example, Havaei et al. 234
[56] created a structure for interactive brain tumor segmentation and applied it to MICCAI-BRATS 235
2013. This study suggested a semi-automatic method to enhance the effectiveness of various 236
classification techniques, including SVM, KNN, and random forests. The improved KNN reported 237
85%, 91%, and 87% for accuracy, specificity, and sensitivity rates, respectively. But SVM 238
outperformed and yielded better rates. 239
Çınarer et al. [57] investigated the statistical features of the input images in order to categorize 240
the data. Then, a variety of methods were employed to examine the accuracy rate of various ML 241
algorithms, including KNN, LDA (linear discriminant analysis), RF (random forest), and SVM. 242
However, KNN’s accuracy rate was lower than SVM's 90% accuracy rate. Kumar et al. [58] 243
proposed a four-modular framework that employed an adaptive KNN classifier to categorize MRI 244
images of brain tumors into normal or abnormal categories. Then, the optimal probabilistic Fuzzy 245
C-Means (FCM) clustering approach is applied to separate the tumor areas. The application of the 246
proposed work on BRATS MICCAI dataset indicated the evaluation's findings show that the 247
suggested technique, which employs KNN-based brain tumor classification, achieved the highest 248
accuracy of 96.5%. The maximum sensitivity was 100%, while the maximum specificity was 93%. 249
Ajai et al. [59] performed analyses on various pre-processing algorithms that can be utilized to 250
enhance images prior to applying the active contour without performing the technique called edge-251
based segmentation. Moreover, both the linear kernel SVM and KNN classifiers' accuracy were 252
compared. According to the findings, KNN is superior to linear SVM for classifying brain tumors 253
when active contouring without an edge-based method of segmentation is utilized. 254
In order to identify and categorize the different types of tumors, Ramdlon et al. [60] created a 255
tumor classification system based on the KNN method that can identify tumors and edema in T1 256
and T2 imaging sequences. The tumor categorization reported 89.5% in terms of accuracy, which 257
can give more precise and detailed information about tumor detection. For the classification of 258
glioblastoma, Wibowo et al. [61] compared the use of KNN and SVM methods. Additionally, the 259
Genetic Algorithm (GA) was applied to identify the chosen relevant features and categorize them 260
using KNN and SVM techniques. With 92.35 percent accuracy, the findings demonstrated that the 261
SVM-GA method outperformed the KNN-GA strategy. 262
Although the KNN technique is much faster than other strategies that require training (such as 263
decision tree, random forest, and SVM), but It does not learn anything in the training period (Lazy 264
Learner). Moreover, as the KNN technique doesn’t need any training dataset before making 265
estimation, new input training samples can be added without impacting the performance of the 266
model. 267
4.1.2 Support Vector Machines 268
SVMs are effective and adaptable supervised ML algorithms used for both classification and 269
regression. However, they are typically implemented for classification. An SVM model is just a 270
hyperplane in a multidimensional space that represents two or more classes. SVM will construct 271
the hyperplane in an iterative manner in order to reduce error [62]. An SVM technique aims to 272
classify samples in order to identify a Maximum Marginal Hyperplane (MMH). As a result of its 273
flexibility in handling several continuous and categorical variables, SVM methods are widely-used 274
ML techniques [63], especially for brain cancer segmentation and classification. For example, 275
Amin et al. [64] developed an automated approach to identify brain tumors whether malignant or 276
benign in MRI images. To compare the proposed framework's precision, an SVM classifier was 277
applied with various cross-validations on the features set. The approach obtained average 278
accuracy, sensitivity, and specificity of 97.1%, 91.9%, and 98.0%, respectively. 279
Padlia et al. [65] suggested a method for identifying and segmenting brain tumors using T1-280
weighted and FLAIR brain images. In order to enhance images and remove noise, a fractional 281
Sobel filter is utilized. Bhattacharya coefficients and mutual information are employed for 282
detecting asymmetry in brain image. After extracting features of the region of interest through the 283
windows and patches, SVM is used to categorize the statistical features to separate the tumor area 284
from the tumor hemisphere. Their method gained an average accuracy of 98.03%. In another study, 285
Khairandish et al. [66] presented a hybrid model, which integrated CNN and SVM models for 286
classification. Also, a threshold-based algorithm for detecting the brain tumor. The hybrid CNN-287
SVM got a 98.49% overall accuracy. Rao and Karunakara [67] focused on efficient classification 288
and segmentation, using KSVM-SSD for more accurate classification. In this study, the malignant 289
tumor is further graded as a low, medium, and high utilizing the SSD (Social Ski Driver) 290
optimization method after being diagnosed as cancerous and non-cancerous using Kernel Support 291
Vector Machine (KSVM). The proposed KSVM-SSD model is shown to be superior regarding 292
classification accuracy assessed on the BRATS datasets, with accuracy values of 99.2%, 99.36%, 293
and 99.15% for the corresponding years of 2018, 2019, and 2020 BRATS datasets. 294
Rashid et al. [68] aimed to shed some light on the area of the brain being damaged by the tumor. 295
The major steps of the described technique include using abnormal MRI brain images as input, 296
anisotropic filtering to remove noise, an SVM classifier for segmentation, and morphological 297
procedures to distinguish the damaged area from the normal one. The results indicated that the 298
segmentation accuracy of the SVM was 83%. For the categorization of medical images, Deepak 299
& Ameer [69] applied CNN features with SVM to enhance the quality of classification. In 300
comparison to the most recent technique, the proposed model outperformed with 95.82% accuracy. 301
Although SVM is able to work well with even semi structured and unstructured data using an 302
appropriate kernel function, but finding a good kernel function is not an easy task. 303
4.1.3 Random Forest 304
Random Forest (RF) strategies are a collection of ML algorithms paired with several classifier 305
trees. Each classifier tree casts a unit vote for the most popular class, and the final sort of result is 306
obtained by combining these results [70]. High classification accuracy, good noise and outlier 307
tolerance, and no overfitting are all features of RFs. In addition to building each tree from a 308
separate bootstrap sample of the data, the RF technique modifies how classification or regression 309
trees are built [71]. In an RF model, node splitting is based on a random subset of features for each 310
tree. By comparing this counterintuitive approach to many other classifiers, it turns out to perform 311
very well. Therefore, this technique has been widely employed for brain cancer tumors. For 312
example, Lefkovits et al. [72] developed and fine-tuned a discriminative RF model to segment 313
brain tumors in multimodal MRI images. Finding the optimal parameter values and the 314
discriminative model's most important constraints is the goal of tuning. The proposed method 315
obtained 75-91% for the whole tumor and 71-82 % for the core section in terms of dice index. 316
Ellwaa et al. [73] expanded a previously described Random Decision Forest (RDF) based brain 317
tumor segmentation technique. The RDF is trained via an iterative process, where certain patients 318
were introduced to the training data using heuristics approaches rather than a randomly selected 319
training dataset. The obtained dice score of the model was reported to be over 80%. 320
Anitha & Raja. [74] introduced brain tumor identification and segmentation methods based 321
on random forest classifiers to classify the brain modality into two groups of normal and abnormal. 322
The proposed approach's sensitivity and specificity rates were 97% and 98%, respectively. Yang 323
et al. [75] utilized Small Kernels of Two-Path Convolutional Neural Network (SK-TPCNN) and 324
RF to provide an automated segmentation approach. The sensitivity scores for the whole tumor, 325
core tumor, and enhancing tumor were 96%, 92.2%, and 83.2%, respectively. Rajagopal. (2019) 326
suggested an approach based on a random forest classifier to detect the Glioma brain tumor. In this 327
study, the characteristics of texture are extracted from MRI images of the brain and optimized 328
using Ant Colony Optimization (ACO) algorithm. The resulting optimized sets of characteristics 329
were categorized thanks to the random forest classification method. The result of this study 330
indicated 97.7% of sensitivity, 96.5% of specificity, and 98.01% of accuracy [76]. 331
Although RF helps to improve the accuracy by diminishes overfitting in decision trees and 332
works well with both continuous and categorical data, but it requires much time for training. 333
4.1.4 Decision Tree 334
Decision tree models (DTs), i.e., non-parametric supervised learning techniques are used in 335
regression and classification. The objective is to learn simple decision rules derived from the 336
features of the data to build a model to predict a target variable’s value. One can assume a tree as 337
a piecewise constant approximation [77], [78]. The tree is utilized to let individuals or 338
organizations evaluate probable actions against each other based on their benefits, costs, and 339
probabilities. A single node is the starting point of a decision tree, which then forks into probable 340
outcomes. Then, every single outcome results in additional nodes, which lead to other probabilities 341
[79]. Previous studies have employed DT approaches for brain cancer diagnosis. For example, 342
Naik and Patel [80] applied a Decision Tree algorithm to discern brain tumors in MRI images and 343
compared the results with a Naive Bayes classification algorithm. The researchers reported that 344
the Decision Tree classifier outperformed a Naive Bayes in the task with an accuracy of 96% and 345
sensitivity of 93%. Chaddad et al. [81] implemented Gaussian Mixture Model (GMM) to take out 346
attributes from brain tumors’ MRI images and applied the Decision Tree classifier to GMM 347
features to evaluate the performance of cancer detection. They defined the task to detect brain 348
tumors based on the T1, T2, and FLAIR MRI images. For the T1 and T2 weighted images, the 349
accuracy performance was 100 %. The accuracy decreased to 94.11 % in FLAIR mode. Hussain 350
et al. [82] implemented multiple feature extraction strategies on MRI images of the brain and so 351
tested the performance of different classification algorithms for tumor detection. The researchers 352
reported that the Decision Tree classifier was the second-best algorithm for the task with a total 353
accuracy of 97.81%, after the Naïve Bayes classifier with total accuracy of 100%. 354
Thayumanavan and Ramasamy [83] developed a framework for the diagnosis and segmentation 355
of tumors in the brain using MRI images and compared the performance of Decision Tree (DT), 356
Random Forest Classifier (RFC), and SVM. The experimental results showed that RFC reached 357
the best result with an accuracy of 98.37%. Moreover, RFC showed a specificity of 99.09%, 358
followed by DT and SVM with 95.68% and 88.78% respectively. Rajendran and Madheswaran 359
[84] developed a hybrid method based on Association Rule Mining and Decision Tree algorithms 360
for brain tumor classification using CT scan images. They showed that the proposed method 361
reached the accuracy and sensitivity of 95% and 97%, respectively. 362
Although DT doesn’t need scaling and normalization of data, but it requires much time for 363
training. Also, a small change in input data leads to large changing in the structure of the model. 364
4.1.5 Artificial Neural Network 365
An ANN is a paradigm for information processing that takes its ideas from how biological 366
nervous systems can learn target patterns. Multiple layers of basic processing units known as 367
neurons make up an ANN structure [85]–[87]. The neuron carries out two tasks: gathering inputs 368
and producing an output. Nodes or artificial neurons are connected to other nodes together with 369
their threshold and weight. When the output of a node is higher than a certain threshold value, that 370
node is activated, and the data is sent to the next layer of the network. Otherwise, data won’t be 371
sent to the next layer [88]. ANN has been widely-used for brain cancer. An SVM and an ANN 372
were utilized by Chithambaram and Perumal [89] to create two hybrid ML models (GA-SVM and 373
GA-ANN), which were then evaluated on two different datasets. Virupakshappa & Amarapur [90] 374
developed a technique for classifying data based on an Adaptive Artificial Neural Network 375
(AANN) approach. Utilizing the Whale Optimization Algorithm (WOA), the values of neurons in 376
the adaptive ANN are optimized. The result of classification accuracy was reported at 98%. 377
Virupakshappa et al. [91] suggested an effective tumor segmentation model based on FCM 378
clustering, multiple feature extraction using Gabor wavelets, and an ANN classifier. Based on the 379
result of this study, it is proved that system accuracy levels up to 85%. 380
Although an ANN model is able to process unorganized data, an effective visual analysis, and 381
ability to process in parallel, but it is economically and computationally expensive and needs a 382
long training process. 383
4.1.6 Naïve Bayes 384
A collection of supervised learning algorithms, Naïve Bayes methods, are founded on 385
implementing Bayes’ theorem with the “naive” assumption that each pair of characteristics is 386
conditionally independent given the value of the class variable [92], [93]. Bayes’ Theorem is a 387
straightforward mathematical procedure for conditional computing probabilities. Conditional 388
probability is defined as the possibility of an event occurring due to the occurrence of another 389
event before it (via assumption, supposition, statement, or evidence). The Naive Bayes Classifier 390
(NBC) is among the easiest and most efficient classification algorithms. It helps develop rapid ML 391
models, which can make accurate predictions [94]. There are several researches that applied Naïve 392
Bayes to diagnose brain cancer in medical images. For example, Kaur and Oberoi (2019) applied 393
a Naïve Bayes classifier to discover brain tumors in MRI images. The proposed algorithm showed 394
86% accuracy for brain tumor segmentation [95]. 395
Raju et al. [96] implemented a Bayesian fuzzy clustering algorithm for the segmentation and 396
the Harmony-Crow Search (HCS) optimization algorithm-based multi-SVNN classifier for the 397
classification of brain tumors. The researchers reported that the proposed method reached an 398
accuracy of 93%. Ulku and Camurcu [97] utilized histogram equalization and morphological 399
image processing techniques to develop a computer-aided brain tumor detection method based on 400
MRI images. Six different classification algorithms were tested. Based on the final results, SVM-401
PSO and KNN algorithms reached 100% accuracy. The Decision Tree algorithm showed 98.11% 402
accuracy in negative samples, while Naïve Bayes showed the weakest performance with 83.71% 403
accuracy in negative samples. 404
Although Naïve Bayes methods are appropriate for solving multi-class prediction problems and 405
require much less training data, but they face the ‘zero-frequency problem’ where they assign zero 406
probability to a categorical sample whose class in the test samples wasn’t available in the training 407
samples. 408
4.1.7 Learning Vector Quantization 409
The Learning Vector Quantization (LVQ) is a prototype-based supervised classification 410
algorithm. With the use of piecewise linear decision surfaces, this algorithm aims to approximate 411
the theoretical Bayes decision boundaries in the input domain of the principal observation vectors 412
[98]. The class codebook vectors are supposedly placed in signal space in an ideal manner. As the 413
classification decision is based on the nearest-neighbor selection among the codebook vectors, its 414
computation is very fast. LVQ algorithms have things in common with other competitive learning 415
algorithms, including Self-Organizing Maps (SOMs) and c-means [99]. This method has been 416
widely used for analyzing medical images and some researchers employed the LVQ approach for 417
brain tumor detection. For example, Liu et al. [100] employed the LVQ neural network for the 418
brain cancer prognosis. A normal diagnostic accuracy of 85.7% and a glioma diagnosis accuracy 419
of 89.5% were achieved by using the suggested procedure. Sonavane et al. [101] proposed a proper 420
and precise classification system for brain tumor detection. This technique uses the LVQ approach 421
to group brain tumors into two groups of normal or abnormal. With respect to the DDSM 422
mammography database and the clinical brain MRI database, the suggested system's accuracy was 423
68.85% and 79.35%, respectively. Sonavane & Sonavane [102] developed a four-stage system and 424
classified brain tumors using the Adaboost technique and LNQ neural network. For clinical brain 425
MRI images, the accuracy rate was 95%, while it was 79.3% for DDSM. 426
The LVQ is intuitive, simple, and easy to implement while still yielding decent performance. 427
However, if the data have a lot of dimensions or are noisy, the Euclidean distance is not a good 428
solution. 429
4.2 Unsupervised Learning 430
Unsupervised learning, as known as unsupervised ML, analyzes and groups unlabeled datasets 431
using ML algorithms. These algorithms cluster data and discover hidden patterns without the 432
intervention of a human [103]–[105]. In other words, unsupervised learning algorithms work based 433
on finding similar attributes, naturally occurring patterns and trends, or relationships in a given 434
dataset [106], [107]. In contrast to supervised learning, it is not possible to apply unsupervised 435
learning methods directly to a regression or a classification problem because there is no knowledge 436
about the values of the output [108]. Unsupervised learning algorithms enable to do more complex 437
processing tasks than supervised learning does. Moreover, dimensionality can be easily reduced 438
thanks to this approach. Unsupervised learning can help to understand raw data. This learning 439
resembles human intelligence as it happens gradually and weighs the result afterwards. Clustering 440
is an unsupervised method that clusters the data into a number of groups. There are many clustering 441
approaches that have been widely used in the medical imaging area, especially clustering brain 442
tumors into a benign and malignant one. 443
Unsupervised learning methods have less complexity in comparison with supervised learning 444
and don’t require to labeled data. However, the outcomes often have lesser accuracy and may be 445
difficult to understand or unpredictable. 446
4.2.1 K-means 447
As a clustering method, K-means is a simple and well-liked algorithm of unsupervised ML 448
[109]. K-means clustering algorithm is prototype-based and seeks to find 𝐾 non-overlapping 449
clusters. This approach aims to categorize a given collection of data into 𝐾 distinct clusters, where 450
k has a predetermined value [110]. Benefiting from linear time complexity, K-means algorithm is 451
optimal for large datasets. K-means uses big unlabeled data to provide deep insights and be 452
beneficial. K-means have been employed in many investigations for brain tumor diagnoses. For 453
example, Khilkhal et al. [111] applied K-means clustering, thresholding, and morphological 454
operations, for segmenting brain tumors in MRI images. The morphological procedure removed 455
non-brain tissue to increase the final accuracy results. The experiments were implemented on 456
BRATS datasets utilizing High-Grade Glioma (HGG) and LGG images. Islam et al. [112] 457
proposed an improved outline to detect brain tumors that makes use of Template-based K-means 458
(TK) with a superpixel technique and Principal Component Analysis (PCA) in order to detect brain 459
tumors efficiently. Their method could obtain an acceptable segmentation result with a shorter 460
processing time in MRI images. Kumar et al. [113] proposed a five stages methodology to segment 461
brain tumors in MRI image segmentation. A rough K-means algorithm was used to achieve this 462
aim. The results are indicative of the fact that the suggested methodology gained better scores in 463
evaluation in comparison with previous works. 464
The K-means model is simple, easy to implement, and Guarantees convergence. However, it is 465
highly dependent on initial values and clustering outliers. 466
467
4.2.2 Fuzzy C-Means 468
Another well-known technique to cluster data is FCM where datasets are classified into 𝑛 469
clusters and is frequently used for pattern recognition [114]. Every segmented data point in the 470
dataset belongs to over one group with distinctive values for membership. If the data is closer to 471
the cluster center, its membership inclines more toward that particular cluster center [115]. 472
Enabling memberships of data points through time may be the main benefit of FCMs clustering. 473
These data points are known to have degrees in [0,1] and this technique makes it possible to show 474
that data points do not necessarily belong to one cluster. Moreover, FCM can provide the best 475
result for the overlapped data set [116]. Devi et al. [117] developed a clustering method called 476
Adaptive Kernel Fuzzy C-Means (AKFCM) for diagnosing brain cancer based on MRI images. 477
Then, they applied a Hybrid Convolution Neural Network- Long Short-Term Memory (CNN-478
LSTM) to enhance the accuracy of tumor categorization. The results indicated that the proposed 479
method did better than the present methods. 480
Debnath et al. [118] applied the Contrast Enhanced Fuzzy C-Means (CEFCM) clustering 481
method to accurately categorize the 2D tumor regions from MR images. Pixel-based Voxel 482
Mapping Technique (PBVMT) mapped the decision values for each pixel location from the 483
segmented image into 3D space and showed the overall accuracy, sensitivity, and specificity of 484
94.8%, 92.14%, and 96.97%, respectively. Sheela et al. [119] proposed a method to segment brain 485
tumors in MRI images based on rotating triangular sections with FCM optimization. At first, the 486
background should mostly be eliminated through morphological reconstruction processes in two 487
levels, after which thresholding happens. To assess the proposed structure’s performance, they 488
employed T1-a weighted contrast-enhanced image dataset. The final assessment of the proposed 489
method is indicated. Soleymanifard et al. [120] applied a classification method, which was a 490
Multiscale Fuzzy C-Means (MsFCM) with 12 scales to discern enhancing tumors in the image 491
cropped from the former stage. Then, thanks to using a neural network classifier, they managed to 492
specify the input tumors’ grade in the MR image. The DICE score results show that the model 493
proposed is highly competitive in comparison with newer segmentation methods. 494
The Fuzzy C-Means model gains best results for overlapped sample points and comparatively 495
better than k-means strategy. However, it is highly dependent on the predefined number of clusters. 496
4.2.3 Mean-Shift Clustering 497
Mean shift clustering algorithm is a powerful nonparametric technique based on centroid that 498
is proven to be useful in various unsupervised learning use cases [121], [122]. This algorithm 499
works by moving data points toward centroids to become the average of nearby points. Mode-500
seeking algorithm is another name for mean shift clustering. The algorithm’s advantage is that it 501
clusters the data without automatically determining how many clusters there should be based on 502
defined bandwidth [123]. Researchers have implemented mean-shift clustering for the clustering 503
of brain tumors. Vallabhaneni and Rajesh [124] developed a method based on mean-shift 504
clustering, and Gray-level Co-Occurrence Matrix (GLCM) features for automatically diagnosing 505
brain tumors in noise-corrupted images. The researchers implemented Edge Adaptive Total 506
Variation Denoising Technique (EATVD) for preserving the edges in the denoising process of 507
images. The experimental results showed an increased precision in tumor detection in noisy 508
images. Singh et al. [125] implemented modified mean-shift-based FCM segmentation for brain 509
tumor detection in MRI images. The results indicated a high level of efficiency and accuracy for 510
the suggested technique. Kim et al. [126] offered a strategy for reducing file sizes and brain tumor 511
discovery in MRI images using modified K-means and mean-shift clustering techniques. The 512
researchers reported that the proposed method reached a precision of 0.914052 and a recall of 513
0.995641. 514
The Mean-Shift clustering approach is able to define the number of clusters automatically and 515
no problem generated from outliers. However, this method unable to work well in case of changing 516
the number of clusters changes abruptly (high dimension). 517
4.2.4 Hierarchical clustering 518
Hierarchical clustering, also known as hierarchical cluster analysis, is an algorithm to categorize 519
similar objects in groups called clusters [127]. Hierarchical clustering includes a multilevel 520
hierarchy tree where clusters at one level are joined as clusters at the next level. The algorithm 521
clusters together object with similar attributes [128]. Finally, the algorithm returns a set of clusters 522
or groups, where clusters are different from each other and objects within clusters are similar to 523
each other. These algorithms have been applied in the medical domain for brain tumor clustering. 524
For instance, Hiratsuka et al. [129] examined samples of human brain tumors using imaging mass 525
spectrometry (IMS). IMS analysis was integrated with the Region Of Interest Analysis (IMS-ROI) 526
and a novel Hierarchical Cluster Analysis (IMS-HCA). IMS-HCA and IMS-ROI appear to be 527
promising methods for finding biomarkers in brain cancer samples, according to the study's 528
findings. Tamilmani & Sivakumari [130] proposed a novel Association Allotment Hierarchical 529
Clustering (AAHC) approach for early brain cancer detection. The final results revealed that the 530
suggested approach reached an accuracy of 100%, outperforming the state-of-the-art method. 531
Moreover, the method was proven to be more computationally efficient compared to other models. 532
The Hierarchical clustering approach is able to define the number of clusters. However, this 533
method unable to work well with huge datasets or vast amounts of data. 534
4.2.5 DBSCAN 535
DBSCAN or Density-Based Spatial Clustering of Applications with Noise is a state-of-the-art 536
algorithm based on density [131]. The prominent feature of this algorithm is detecting clusters in 537
all sizes and shapes in different databases, even those with noise and outliers [132]. This approach 538
was employed to diagnose brain tumors. For instance, Muthaiyan et al. [133] proposed systems to 539
discover brain tumors using images taken using MRI and PET by different classification methods, 540
such as Gustafson-Kessel (G-K) algorithm, k-means clustering algorithm, DBSCAN, and FCM. 541
The effectiveness of various algorithms is compared, and DBSCAN is not among the most 542
effective methods for detecting brain cancer. Moreover, Bandyopadhyay [134] used DBSCAN and 543
K-means clustering for the problem of segmenting and grouping brain tumors from MRI images 544
of the human brain, and the result of the study indicated the effectiveness of the DBSCAN 545
approach in comparison with the K-means technique. 546
The DBSCAN clustering approach not only is able to define the number of clusters but also is able 547
to find arbitrarily shaped and size of clusters. However, this method fails in case of varying density 548
clusters. 549
4.2.6 Gaussian Mixture Models 550
As a probabilistic model, GMM presumes that a combination of a limited number of Gaussian 551
distributions, which have unknown parameters, generate all the data points [135]. To investigate 552
the effectiveness of GMM in brain cancer segmentation, Chaddad [136] introduced a unique 553
technique for extracting Glioblastoma (GBM) features from MRI data using GMM. The accuracy 554
performance for the T1-WI and T2-WI was 97.05% and 97.05%, respectively. The accuracy 555
dropped to 94.11 % in FLAIR mode. These experimental findings show promise for improving 556
heterogeneity features and, consequently, early GBM treatment. Pravitasari et al. [137] segmented 557
the MRI-based brain tumor utilizing the GMM and the Reversible Jump Markov Chain Monte 558
Carlo algorithm. The study's findings showed that the suggested technique executed the algorithm 559
quickly and efficiently. 560
The GMM approach is less sensitive to the number of parameters. However, this method has a 561
slow convergence rate and is sensitive to initialization values. 562
4.3 Overview of Supervised and Unsupervised Methods Application 563
Table 6 displays the 12 papers that applied either supervised or unsupervised approaches. It 564
should be mentioned that all these papers are selected randomly. According to this table, as most 565
of the researchers applied the BRaTS dataset, MRI image modalities are T1, T1c, T2, and FLAIR. 566
The advantages of each approach are mentioned in the last column. 567
568
Table 6. Overview of recent segmentation methods. 569
Author Method of
Segmentation
MRI Modalities Dataset Advantage
Bonte et al.[138] Random Forest T1ce and FLAIR
MRI BRaTS 2013 database
1. Requires fresh
training sets.
2. Very much
sensitive to noise
Bahadure et al.
[139]
watershed
-
15 images with 9 slices
using 3 Tesla Siemens
Magnetom Spectra MR
machine
1. Sensitivity to
intensity variants.
2. Ignore the blurry
boundaries.
3. High system
complexity.
FCM
DCT
BWT
Author Method of
Segmentation
MRI Modalities Dataset Advantage
Reddy & Reddy.
[140]
Region growing
algorithm
T1, T1c, T2, and
FLAIR BRATS 2015
1. Over
segmentation.
2. Need seed point
selection.
Xie & Xiaozhen.
[141] KNN T1 and T2 -
1. Poor run-time
performance.
2. Only using T1 and
T2 modalities.
Raja & Rani.
[142]
Bayesian fuzzy
clustering
T1, T1c, T2, and
FLAIR BRATS 2015
1. High system
complexity.
2. Time
-
consuming.
Ilhan & Ilhan.
[143]
Novel Threshold-
based method Flair, T2 and T1C TCIA: 100 MRI images
1. Poor run-time
performance.
2. High system
complexity
Aslamet al. [144] Sobol Edge-
detection - -
1. Not acceptable
result at fuzzy
borders.
2. Not clear the
process of the
model.
Kermi et al. [145] region-based +
boundary
-
based
T1, T1ce, T2, and
Flair
BRaTS'2017 1. Poor run-time
performance
Sheela &
Suganthi. [146]
Region of Interest +
Region Growing +
Morphological
Operation
- medical MR Images of
120 patient
1. High system
complexity.
2. Class imbalances
not considered.
Khan et al. [147] k-means+ VGG19 T1, T1c, T2, and
FLAIR BRaTS 2015
1. Many features are
required.
2. Easy to lose
information
ŞİŞİK & SERT.
[148] BTS-ELM-FRFCM T1 3200 pieces MRI image
1. Only using T1
modality.
2. Class imbalances
not considered.
Khosravanian et
al. [149]
superpixel fuzzy
clustering +lattice
Boltzmann
T1, T1ce, T2, and
Flair BRaTS 2017
1. Ignore the blurry
boundaries.
2. High system
complexity
570
5. Deep Learning Application in Brain Tumor Segmentation 571
572
DL is a form of ML and AI that mimics how humans gain specific subjects. In data science, 573
which also encompasses statistics and predictive modeling, DL plays a significant role. Multiple 574
visual analysis tasks, including classification, object detection, and tracking, have provided notable 575
performance gains due to DL techniques [150]–[154]. Additionally, DL techniques have had a 576
significant impact on the automation of medical image processing tasks while showing state-of-577
the-art accuracy. Over the last decades, many DL techniques have been developed and applied in 578
different areas, including brain tumor diagnosis [155]. In the following, the popular DL approaches 579
that are utilized for brain cancer diagnosis are introduced. Fig. 5 illustrates the number of papers 580
that applied different approaches of DL to investigate their performance in brain cancer 581
segmentation and classification. According to the Fig. 5, the number of studies that applied CNN-582
based approaches is considerably higher than other DL approaches with 1793 publications between 583
2015 to 2022. LSTM is the second DL approach that has been employed in brain tumor diagnosis 584
with 282 papers. 585
586
587
Fig. 5. Number of publications between 2015 to 2022 for brain tumor segmentation using DL models. 588
589
5.1 Convolutional Neural Network 590
A Convolutional Neural Network (CNN), or ConvNet, sets the basis for DL [156], [157]. In 591
this method, learning is received straight from the data, so there is no need to extract features 592
manually [51]. The most significant usage of CNNs is in discovering patterns in images [158]. A 593
CNN network can be trained using a large dataset from scratch by fine-tuning an existing model 594
or utilizing "off-the-shelf CNN features” [152]. Fine-tuning involves transferring weights of the 595
first 𝑛 layers learned from an earlier-based network to the new network [159]. The dataset obtained 596
for the new network is trained to perform specific tasks. By effectively learning general image 597
features through transfer learning, CNNs are able to tackle the majority of computer vision 598
problems by combining these features with straightforward classifiers [50]. This is the main reason 599
why the CNN approach has been applied considerably for brain tumor diagnosis. For example, 600
Abd El Kader et al. [160] used deep differential CNN to categorize brain tumors in MRI images. 601
This method achieved maximum accuracy of 99.25%. 602
Bacanin et al. [161] deployed the firefly algorithm to optimize CNN for glioma brain tumor 603
grade classification. The researchers reported that the introduced method reached the maximum 604
multi-class accuracy of 97.9% and the maximum accuracy of 96.5% in the images containing brain 605
tumors. Wang et al. [162] offered a ground-breaking structure based on CNN with a dilated 606
convolutional feature pyramid called DFP-ResUNet to categorize multimodal brain tumors. The 607
results of testing the suggested model on the BRaTS 2018 dataset showed the mean Dice value of 608
different subregions to be Enhancing Tumor (ET) 0.8431, Whole Tumor (WT) 0.897, and Tumor 609
Core (TC) 0.9068. Gurunathan and Krishnan [163] applied CNN for brain tumor detection and 610
diagnosis using MRI images. The method reached the maximum classification accuracy of 98.3%, 611
the sensitivity of 97.2%, and the specificity of 98.9%. 612
The CNN models are able to explore hidden patterns inside the input data automatically and 613
share weights between layers. However, these models fail to encode the orientation and position 614
of objects. 615
616
5.2 Recurrent Neural Network 617
As an ANN, a Recurrent Neural Network (RNN) employs time series data or sequential data 618
[164]. RNNs have the concept of "memory," which enables the method to retain the qualities or 619
details of the former inputs in order to produce the following output in the sequence. RNNs are 620
characterized by their capacity to transmit data over time steps [165]. RNNs feature an extra 621
parameter matrix in their structure for connections between time steps, which encourages training 622
in the temporal domain and makes use of the input's sequential nature. In the RNN technique, the 623
predictions made at each time step are trained to be based on both the most recent input and data 624
from earlier time steps [166]. RNNs hold the second place of the most favored approaches to 625
diagnose brain cancer and have been implemented in many studies. For example, SivaSai et al. 626
[167] used fuzzy RNN to categorize brain tumors in MRI images automatically. The results 627
showed an accuracy of 87.8% and it was proven that the proposed framework is much more 628
computationally efficient. Zhou et al. [168] deployed DenseNet and RNN for a holistic screening 629
and categorization of brain tumors via MRI images. By testing the proposed structure on public 630
datasets, the researchers reported that the DenseNet-RNN approach reached the maximum 631
accuracy of 84.61%. However, DenseNet-LSTM achieved an accuracy of 92.13% and performed 632
better than RNN based approach. Begum and Lakshmi [169] combined optimal wavelet statistical 633
texture and RNN to discover and classify brain tumors in MRI images. The proposed method 634
achieved a maximum accuracy of 96% and a maximum sensitivity of 100% in the classification. 635
Moreover, the introduced approach reached the maximum accuracy and sensitivity of 95% and 636
97%, respectively for the segmentation. 637
The RNN models are capable of processing inputs of any length and each sample can be 638
assumed to be dependent on former samples. However, these models face issues like Vanishing 639
Gradient or Exploding Gradient. 640
641
5.3 Long Short-Term Memory 642
There are recurring neural networks able to learn order dependency in issues related to 643
predicting sequences; these networks are called Long Short-Term Memory (LSTM) networks 644
[170]. LSTM is one of the most widely-used RNN designs to date. It is the best option for modeling 645
sequential data and is thus utilized to learn the complex dynamics of human behavior. The word 646
"cell state" refers to long-term memory. Previous data is stored in the cells because of their 647
recursive nature. LSTM was specifically created and developed in order to address the 648
disappearing gradient and exploding gradient issues in long-term training [171]. Fig. 6 shows an 649
example of LSTM structure and the way this technique works. 650
651
Fig. 6. An example of LSTM structure. 652
653
This approach has been utilized in previous studies to diagnose a brain tumor. For example, 654
Dandıl and Karaca [172] used stacked LSTM for pseudo brain tumor detection based on MRI 655
spectroscopy signals. The experimental results indicated an accuracy of 93.44% for the 656
categorization of pseudo brain tumor with glioblastoma, 85.56% accuracy for a pseudo brain tumor 657
with diffuse, 88.33% for a pseudo brain tumor with astrocytoma, and 99.23% for a pseudo brain 658
tumor with metastatic brain tumors. Xu et al. [173] proposed an LSTM Multi-modal UNet to 659
categorize tumors using multi-modal MRI. The experimental results by testing the model's 660
performance on the BRATS-2015 dataset showed that the proposed LSTM multi-modal UNet 661
outperformed the standard U-Net with fewer model parameters. Shahzadi et al. [174] developed an 662
approach according to a cascade of CNN with an LSTM network for 3D brain tumor MR image 663
classification. The researchers reported an accuracy of 84% for the proposed method. 664
The LSTM models are capable of learning long-term dependencies. However, these models 665
prone to overfitting and need a lot of resources, high memory-bandwidth, and time to get trained. 666
667
5.4 Generative Adversarial Network 668
Generative Adversarial Network (GAN) is one of the categories of generative models [175]. 669
These models have the capacity to create or develop new data with the statistics similar to the 670
training set given. For instance, a GAN trained on images can produce new images with numerous 671
realistic features that appear to be created by humans, at least on the surface [176]. Fig. 7 indicates 672
the simple structure of the GAN approach. 673
674
Fig. 7. An example of the GAN method structure. 675
676
In order to evaluate the efficiency of the GAN method in segmenting brain tumors, several 677
studies have been conducted. For example, Nema et al. [177] developed a novel architecture called 678
residual cyclic unpaired encoder-decoder network (RescueNet) to segment brain tumors which is 679
trained based on an unpaired GAN. The results showed a Dice value of 0.9401% and 0.9463% for 680
the BRaTS 2015 and BRaTS 2017 datasets, respectively. Neelima et al. [178] provided an 681
approach founded on the Optimal DeepMRSeg strategy trained by a devised Sailfish Political 682
Optimizer (SPO) algorithm to segment the tumors and then applied GAN to classify brain tumors 683
using MRI images. The elevated accuracy, segmentation accuracy, sensitivity, and specificity 684
resulted from this method were 91.7%, 90%, 92.8%, and 92.5%, respectively. Rezaei et al. [179] 685
proposed a new method according to the adversarial network called voxel-GAN for mitigating 686
imbalanced data problems in segmenting the tumors using 3D brain MR or CT images. Once the 687
suggested method was evaluated on the ISLES dataset, the results showed a Dice value of 0.83, a 688
Hausdorff score of 9.3, a precision of 0.81, and a recall of 0.78. 689
The LSTM models are capable of improving data instances, lowering costs, and increasing 690
data production. However, these models require strong technical knowledge and advanced 691
datasets. 692
693
5.5 Deep Belief Network 694
Deep Belief Networks (DBNs) are a type of graphical representation that is fundamentally 695
generative, producing all possible values for the given case. It combines ML and neural networks 696
with probability and statistics [180]. DBNs are made up of many layers with values, but there is 697
no relationship between the values and the layers. The major objective is to assist the system in 698
categorizing the data into distinct categories. Fig. 8 represents a simple structure of the DBN 699
approach. 700
701
Fig. 8. An example of DBN method structure. 702
703
In other cancer studies, it is indicated that DBN is among the best approaches to segment 704
tumors. There are several research works that used DBN for brain tumor segmentation. For 705
example, Kharrat and Néji [181] implemented personalized DBNs for brain tumor classification 706
using MRI images. The researchers trained the proposed model on the BRaTS dataset and showed 707
the model reached an accuracy of 91.6%. Raju et al. [182] offered a strategy according to a DBN 708
and hybrid active contour model to split and classify the tumors of the brain via MRI images. The 709
outcomes indicated the accuracy of 0.945, sensitivity of 0.9695, and specificity of 0.99348 for the 710
proposed model. 711
The DBN models are capable of providing the best performance results even if the amount of 712
data is huge. However, these models require strong technical knowledge and huge data to perform 713
better. 714
715
5.6 Autoencoders 716
Autoencoders are an unsupervised learning method that uses neural networks for the 717
presentation learning task [183]. Specifically, it is a neural network architecture with an imposed 718
bottleneck that represents a compact knowledge of the specified input [184]. Amin et al. [185] 719
deployed Stacked Sparse AutoEncoders (SSAE) to detect brain tumors. They tested the suggested 720
model on the BRaTS datasets and reached the average accuracies of 100%, 90%, 95%, 100%, 721
97%, and 95% on the 2012 dataset, 2012 synthetic dataset, 2013 dataset, Leaderboard 2013 dataset, 722
2014 and 2015 datasets, respectively. Badža and Barjaktarović [186] implemented a convolutional 723
autoencoder to segment brain tumors based on MRI images. The researchers reported that the 724
proposed method achieved 99.23% average accuracy for pixel classification and an average 725
accuracy of 99.28% for 5-fold cross-validation and one test. 726
The Autoencoders can reduce the dimensionality of the data, provide an appropriate way to 727
diminish the noise of input data greatly, and make the creation of DL frameworks much more 728
efficient. However, learning and reproducing input features in training autoencoders is unique to 729
the data they are trained on and don't work for new data. 730
731
5.7 Reinforcement Learning 732
Reinforcement Learning (RL) is the training of ML models to make a sequence of decisions. 733
The agent gains the ability to do a task in a possibly complex and uncertain environment. An AI 734
encounters a scenario similar to a game during RL. In order to solve the problem, the computer 735
uses trial and error. AI is rewarded or punished for the steps it takes to make the machine do what 736
the programmer desires. The goal of the RL algorithm is to maximize the total reward Stember 737
and Shalu [187] automatically extracted labels from clinical reports and then utilized deep RL 738
classifier to classify3D MRI brain volumes. The results revealed that the presented method 739
provides 100% accuracy. 740
The RL models are desired to gain long-term outcomes, which are challenging to obtain. 741
Moreover, a perfect model can be created to solve a particular problem. However, RL models need 742
plenty of data and computation and are not preferable to utilize for solving simple problems. 743
744
5.8 Gate Recurrent Unit 745
Gated Recurrent Unit (GRU) is the upgraded version of the standard RNN, i.e. RNN, which 746
was offered by Kyunghyun Cho et al. [188] GRUs are very similar to Long Short Term Memory 747
(LSTM). Resembling the LSTM, GRU employs gates to supervise the flow of information, which 748
are newer than LSTM. Therefore, they improve LSTM and focus on simplifying the architecture. 749
Fig. 9 displays the structure of the GRU model. 750
751
Fig. 9. An example of GRU method structure. 752
753
One of the main usages of the GRU is increasing the memory capacity of a RNN and facilitating 754
the training of a model. On top of that, we can employ the hidden unit and solve the problem of 755
vanishing gradients in RNNs. There are investigations that applied the GRU technique for brain 756
cancer diagnosis. For example, Gab Allah et al. [189] employed a combination of a VGG-16 757
network and a GRU model to increase the brain tumor segmentation in the presence of even fuzzy 758
borders. Table 7 indicates the 16 papers that used DL approaches to classify and segment brain 759
tumors. 760
Unlike LSTM, GRU models do not have a separate cell state (Ct) and only have a hidden state 761
(Ht). These models are much faster due to the simpler architecture. 762
763
Table 7. Overview of segmentation through DL approaches. 764
Author Method of
Segmentation
MRI
Modalities Dataset Limitations
Chang et al. [190] CNN+FCRE T1, T2, T1c,
FLAIR BRaTS2013
1. Requires fresh training
sets.
2. Only using HGG patient
subjects.
Kaldera et al. [191] Faster R-CNN T1
MRI dataset of
Nanfang Hospital
and General
Hospital in China
1. Requires fresh training
sets.
2. Only using T1 modality.
Sajjad et al [192] CNN + Data
augmentation T1 Radiopaedia
dataset
1. Need a post-processing
step.
2. Only using T1 modality.
Murthy et al [193] Optimized CNN - kaggle
1. Requires fresh training
sets.
2. Need a post-processing
step.
Rehman et al [194] 3D CNN T1, T2, T1c,
FLAIR
BRaTS 2015,
2017, and 2018
1. Fail to detect the
unobvious and small brain
tumors.
2. High system complexity
Amin et al [195] LSTM T1, T2, T1c,
FLAIR
BRaTS +SISS
database
1. High system complexity.
2. Requires fine-tuning of the
network parameters.
Liu et al [196] DRL + DTCWT T1, T2, T1c,
FLAIR
BRaTS2018 +
CQ500 + hospital
databases
1. High system complexity
2. Lack of spatial
consistency.
Kumar et al [197] DBN + GS-MVO T1 Kaggle dataset
1. The feature selection is not
clear.
2. Only using T1 modality.
Author Method of
Segmentation
MRI
Modalities Dataset Limitations
Harish & Baskar
[198]
R-CNN + Alex Net
model - -
1. Many layers, which
increases system complexity.
2. Requires fresh training
sets.
Ahmad et al. [199] Variational
Autoencoders + GAN T1 Figshare
1. Requires fine-tuning of the
network parameters
2. Only using T1 modality.
Mukherkjee et al.
[200]
Optimized GAN
method
T1, T2, T1c,
FLAIR
BRaTS 2020
dataset
1. High computational cost.
2. High system complexity.
Takrouni & Douik
[201] DPGM+DDM T1, T2, T2c &
FLAIR
BRaTS 2013 +
2015 +2017
1. Need a lot of training data.
2. High system complexity.
Chattopadhyay &
Maitra [202] CNN T1, T2, T2c &
FLAIR BRaTS 2020
1. Weak interpretability.
2. Requires fresh training
sets.
Kesav & Jibukumar
[203] RCNN T1 Figshare + Kaggle
1. Requires fine tuning of the
network parameters
2. Only using T1 modality.
Vankdothu et al
[204] CNN + LSTM T1 Kaggle 1. High system complexity.
2. Only using T1 modality.
Ranjbarzadeh et al.
[18] CNN + Attention T1, T2, T2c &
FLAIR BRaTS 2018
1. High system complexity.
2. Requires fresh training
sets.
765
6. Brain Tumor Imaging Database 766
There are many advanced databases of medical images taken from patients suffering from brain 767
tumors to facilitate the development and validation of new images. Table 8 outlines the list of top 768
brain databases used in recent investigations. The number of data and their modalities mentioned. 769
770
Table 8. Top database for the brain tumors. 771
Dataset From Number of images Available Modalities
BRATS 2012 MICCAI 2012 Challenge 45 3D T1, T2, T1c, FLAIR
BRATS 2013 MICCAI 2013 Challenge 65 3D T1, T2, T1c, FLAIR
BRATS 2014 MICCAI 2014 Challenge 50 3D T1, T2, T1c, FLAIR
BRATS 2015 MICCAI 2015 Challenge 300 3D T1, T2, T1c, FLAIR
BRATS 2016 MICCAI 2016 Challenge 300 3D T1, T2, T1c, FLAIR
BRATS 2017 MICCAI 2017 Challenge 285 3D T1, T2, T1c, FLAIR
BRATS 2018 MICCAI 2018 Challenge 285 3D T1, T2, T1c, FLAIR
BRATS 2019 MICCAI 2019 Challenge 335 3D T1, T2, T1c, FLAIR
BRATS 2020 MICCAI 2020 Challenge - T1, T2, T1c, FLAIR
Dataset From Number of images Available Modalities
BRATS 2021 MICCAI 2021 Challenge - T1, T2, T1c, FLAIR
ADNI1 Alzheimer’s disease neuroimaging
initiative 400 T2, FLAIR, DTI
BrainnWeb McConnell Brain Imaging
Centre 21 T1, T2, PD- weighted
RIDER TCIA 365 3D T1, T2- weighted
AANLIB Harvard Medical School - T1, T2- weighted MRI
The IBSR The CMA 39 T1- weighted
Allen brain atlas Allen Institute Publications
for Brain Science 20 T1, T2, DTI
Figshare Jun Change 3064 T1
Kaggle - 3264 -
SISS - - FLAIR, DWI, T2, T1
772
6.1 BRaTS Database 773
According to Table 8, BRaTS are the most used data set for brain cancer diagnosis. BRaTS 774
database employs special MRI scans that are multi-institutional and pre-operative to concentrate 775
on parting incongruous brain tumors, i.e., gliomas, which are different in appearance, shape, and 776
histology. Table 9 indicates the dice and Hausdorff parameter for the papers that applied the 777
different methods to diagnose brain tumors based on BRaTS database. 778
779
Table 9. Dice and Hausdorff index of investigations based on BRaTS 2020-2017. 780
BRaTS 2020
Method
Dice Hausdorff 95%
Comp Core Enh Comp Core Enh
3D U-Net
[205]
89% 84% 81% 6.4 19.4 15.8
nnU-Net
[206]
89% 85% 82% 8.50 17.33 17.80
Multiple U-net
[207]
89% 84% 78% 6.7 19.55 20.4
Scale Attention Network
[208]
88% 84% 82% 5.2 17.97 13.43
Variational-autoencoder + regularized 3D U-Net
[209]
89% 79% 70% 4.62 10.07 34.30
Deep Layer Aggregation
[210]
88% 83% 79% 5.32 22.32 20.44
Lightweight U-Nets
[211]
87% 80% 75% 6.2 19.6 21.46
HI-Net
[212]
88% 84% 79% - - -
Modified UNet
[213]
89% 85% 79% - - -
DR-Unet104
[214]
87% 80% 75%
10.41 21.84 24.68
BRaTS 2019
Method Dice Precision
Comp Core Enh Comp Core Enh
Two-stage Cascade UNet
[215]
89% 84% 83% 4.62 4.13 2.65
Encoder-decoder + Combined loss function
[216]
88% 84% 83% 4.7 3.97 2.20
CNN
[217]
87% 78% 78% 7.3 6.8 3.7
3D UNet
[218]
85% 80% 78% 6.5 6.3 3.5
RMU-Net
[219]
92% 91% 83% - - -
Deep Layer Aggregation
[220]
87% 83% 79% 5.32 22.32 20.44
DNN
[221]
88% 86% 81% 4.8 5.6 2.4
Multi-Resolution 3D CNN
[222]
82% 72% 70% 8.42 9.14 5.59
3D Residual U-Net
[223]
83% 77% 70% 14.64 26.69 25.56
3D FCN
[224]
89% 78% 76% - - -
BRaTS 2018
Method
Dice Precision
Comp Core Enh Comp Core Enh
3D UNet
[225]
86% 82% 76% 7.01 5.63 5.6
Deep CNN
[226]
88% 79% 78% 5.5 6.9 2.93
3D UNet
[227]
87% 77% 71% 6.5 8.31 4.14
Contour-aware 3D CNN
[228]
89% 79% 72% 8.05 7.5 5.2
S3D-UNet
[229]
84% 78% 69% 9.2 7.7 4.5
HTTU-Net
[230]
88% 89% 82% 7.53 8.81 4.43
Auto-encoder Regularization
[231]
88% 81% 77% 5.9 4.8 3.8
Cascaded UNet
[232]
88% 78% 72% - - -
CNN
[233]
87% 77% 78% 6.55 27.05 15.90
CNN + Test-time Augmentation [234] 88% 80% 75% 5.97 6.71 4.16
BRaTS 2017
Method
Dice Precision
Comp Core Enh Comp Core Enh
UNet
[235]
81% 76% 65% - - -
Random Forest +CNN
[236]
85% 69% 67% 6.12 28.72 23.55
FCN
[237]
70% 55% 40% - - -
Anisotropic CNN
[238]
87% 77% 78% 6.5 27 15.90
EMMA
[239]
89% 80% 73% 5.01 23.1 36.0
CNN
[240]
83% 65% 65% - - -
UNet
[241]
88% 76% 64% - - -
Pixel Net
[242]
87% 76% 68% 9.8 12.30 12.93
Multi-path CNN
[243]
84% 69% 60% - - -
Ensemble-of-Forest
[244]
42% 41% 42% 21.17 40.06 69.12
781
7. Performance measures 782
The performance assessment of the classification or segmentation strategies can be 783
accomplished using different methods. Researchers employ various strategies for validating the 784
obtained outcomes. The most popular and widely used performance measures 785
including Sensitivity (Recall or True Positive Rate), Specificity, Accuracy, and Precision, 786
Confusion matrix, Jaccard Index, and Dice Similarity. These criteria can be defined as follows: 787
Confusion matrix (CM) is used to provide crucial information about actual and estimated 788
outcomes created by the classification or segmentation techniques. One example of a two-class 789
classification task is demonstrated as in Table 10. 790
791
Table 10. The details of classification criteria for two classes. 792
Category Estimated Brain tumor Ordinary tissue
Brain tumor True Positive (TP) False Negative (FN)
Ordinary tissue False positive (FP) True Negative (TN)
793
Where TP,FP, FN, and TN are described as: 794
TP: Correctly classified or segmented brain tumor. 795
TN: Correct classified or segmented of Ordinary tissue as Ordinary tissue. 796
FN: Wrong classified or segmented of actual tumor tissue as Ordinary tissue. 797
FP: Wrong classified or segmented of an ordinary tissue. 798
799
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦
=
𝑇𝑁
+
𝑇𝑃
𝑇𝑁
+
𝑇𝑃
+
𝐹𝑃
+
𝐹𝑁
(1)
𝐷𝐼𝐶𝐸
=
2
×
𝑇𝑃
(
2
×
𝑇𝑃
)
+
𝐹𝑃
+
𝐹𝑁
(2)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛
=
𝑇𝑃
𝑇𝑃
+
𝐹𝑃
(3)
Specificity
=
𝑇𝑁
𝑇𝑁
+
𝐹𝑁
(4)
Sensitivity
or
Recall
=
𝑇𝑃
𝑇𝑃
+
𝐹𝑁
(5)
Jaccard
Index
(
A
,
B
)
=
|
𝐴
∩
𝐵
|
|
𝐴
∪
𝐵
|
(6)
800
8. Discussion 801
The brain tumor is one of the fatal diseases that occurs when the growth of cells in the brain is 802
out of control. The mortality rate of this cancer made researchers investigate approaches for early 803
brain cancer diagnosis. MRI images are one of the best tools to diagnose cancer by providing a 804
picture of soft tissue in the brain. Over the last decades, many ML-based and DL-based approaches 805
have been developed. However, due to the large number of articles that implemented these 806
approaches, it is important to summarize the current studies and methods. In this work, we provide 807
a holistic approach and summarize ML-based segmentation approaches, DL-based segmentation 808
methods, a review of top DL and ML papers, the top database of brain cancer, and a comparison 809
of accuracy rate in applying different methods on publicly available datasets. The application of 810
CAD systems that work based on DL and ML approaches in brain tumor diagnosis increases 811
accuracy, decreases failure in diagnosis, early detection, and provides better treatment approaches. 812
Also, in underdeveloped countries with a scarcity of experts, CAD-based systems can provide 813
early diagnosis, which results in a decrease in mortality rate. 814
The analyses of the previous studies indicated that MRI is the best imaging technique for brain 815
tumor diagnosis (Tables 6 and 7). The main reason for the widespread usage of MRI is that this 816
imaging technique provides more details compared to other approaches like CT scans. Moreover, 817
over the last years, the application of DL approaches was significantly more than ML techniques. 818
According to Figs. 4 and 5, the number of total papers that applied DL approaches has increased 819
significantly. However, the number of papers that implemented ML techniques or hybrid methods 820
is still more than DL-based approaches. SVM and CNN are the most used ML and DL approaches 821
for brain tumor segmentation. Moreover, in most papers, BRaTS datasets are employed. According 822
to Table 9, the dice and Hausdorff rate indicate that the BRaTS dataset can also be used in future 823
studies as they are reliable. 824
825
9. Conclusion 826
Brain tumor segmentation has widely benefitted from advancements in AI. Researchers have 827
been applying AI algorithms and techniques for detecting brain tumors, computing tissue volumes, 828
abnormality detection, pathology, planning of treatments, and computer-aided surgery. These 829
techniques work well in tasks related to segmenting brain tumors as their features enable 830
distinguishing abnormal tissues from normal ones. This paper offers a general survey of methods 831
applied to brain tumor segmentation. A long array of automatic and semi-automatic brain tumor 832
segmentation, classification, and feature extraction methods is covered in this study. The current 833
paper quantitatively measures the up-to-date approaches based on multiple evaluation metrics to 834
help readers and medical experts both to develop future research directions and, more importantly, 835
identify the most effective and precise strategies to segment tumors of the brain. This paper 836
proposes that the upcoming research aiming to enhance the performance of the current systems for 837
brain segmentation can be followed in several directions: (i) Gathering larger databases with 838
images from various qualities, (ii) Focusing on improving the classification accuracy of current 839
methods by developing novel methods for feature extraction, and (iii) Developing hybrid systems 840
consisting of multiple approaches regarding ML and DL. 841
Except the techniques employed in this study, some of the most representative computational 842
intelligence techniques can be utilized to solve the problems, like Harris hawks optimization 843
(HHO), Monarch Butterfly Optimization (MBO), Earthworm Optimization Algorithm (EWA), 844
Moth Search Algorithm (MSA), Colony Predation Algorithm (CPA), Slime Mould Algorithm 845
(SMA), Hunger Games Search (HGS), and Runge Kutta Optimizer (RUN). 846
However, it is still questionable whether the abundance of computational resources, deep learning 847
technology, and training data needed to run DL at full performance is meaningful, considering 848
other learning techniques that may yield fast, higher interpretability, and close performance with 849
less parameterization, tuning, and fewer resources. 850
851
Acknowledgement 852
This publication has emanated from research [conducted with the financial support of/supported 853
in part by a grant from Science Foundation Ireland under Grant number No. 18/CRT/6183 and is 854
supported by the ADAPT Centre for Digital Content Technology which is funded under the SFI 855
Research Centres Programme (Grant 13/RC/2106/_P2), Lero SFI Centre for Software (Grant 856
13/RC/2094/_P2) and is co-funded under the European Regional Development Fund. For the 857
purpose of Open Access, the author has applied a CC BY public copyright license to any Author 858
Accepted Manuscript version arising from this submission. 859
860
Declaration of interests 861
The authors declare that they have no known competing financial interests or personal 862
relationships that could have appeared to influence the work reported in this paper. 863
864
865
866
References 867
[1] A. Petruzzi, C. Y. Finocchiaro, E. Lamperti, and A. Salmaggi, “Living with a brain tumor,” 868
Supportive Care in Cancer 2012 21:4, vol. 21, no. 4, pp. 1105–1111, Oct. 2012, doi: 869
10.1007/S00520-012-1632-3. 870
[2] C. Kruchko, Q. T. Ostrom, H. Gittleman, and J. S. Barnholtz-Sloan, “The CBTRUS story: providing 871
accurate population-based statistics on brain and other central nervous system tumors for everyone,” 872
Neuro Oncol, vol. 20, no. 3, pp. 295–298, Feb. 2018, doi: 10.1093/NEUONC/NOY006. 873
[3] A. Tahir, M. Asif, M. Bin Ahmad, T. Mahmood, M. A. Khan, and M. Ali, “Brain Tumor Detection 874
using Decision-Based Fusion Empowered with Fuzzy Logic,” Math Probl Eng, vol. 2022, pp. 1–875
13, Aug. 2022, doi: 10.1155/2022/2710285. 876
[4] K. D. Miller et al., “Brain and other central nervous system tumor statistics, 2021,” CA Cancer J 877
Clin, vol. 71, no. 5, pp. 381–406, Sep. 2021, doi: 10.3322/CAAC.21693. 878
[5] J. B. Iorgulescu et al., “Molecular biomarker-defined brain tumors: Epidemiology, validity, and 879
completeness in the United States,” Neuro Oncol, Apr. 2022, doi: 10.1093/NEUONC/NOAC113. 880
[6] P. A. McKinney, “Brain tumours: incidence, survival, and aetiology,” J Neurol Neurosurg 881
Psychiatry, vol. 75, no. suppl 2, pp. ii12–ii17, Jun. 2004, doi: 10.1136/JNNP.2004.040741. 882
[7] F. Özyurt, E. Sert, and D. Avcı, “An expert system for brain tumor detection: Fuzzy C-means with 883
super resolution and convolutional neural network with extreme learning machine,” Med 884
Hypotheses, vol. 134, p. 109433, Jan. 2020, doi: 10.1016/J.MEHY.2019.109433. 885
[8] N. Tataei Sarshar et al., “Glioma Brain Tumor Segmentation in Four MRI Modalities Using 886
a Convolutional Neural Network and Based on a Transfer Learning Method,” pp. 386–402, 2023, 887
doi: 10.1007/978-3-031-04435-9_39. 888
[9] N. A. Charles, E. C. Holland, R. Gilbertson, R. Glass, and H. Kettenmann, “The brain tumor 889
microenvironment,” Glia, vol. 59, no. 8, pp. 1169–1180, Aug. 2011, doi: 10.1002/GLIA.21136. 890
[10] B. Cacho-Díaz, D. R. García-Botello, T. Wegman-Ostrosky, G. Reyes-Soto, E. Ortiz-Sánchez, and 891
L. A. Herrera-Montalvo, “Tumor microenvironment differences between primary tumor and brain 892
metastases,” Journal of Translational Medicine 2020 18:1, vol. 18, no. 1, pp. 1–12, Jan. 2020, doi: 893
10.1186/S12967-019-02189-8. 894
[11] I. Koh and P. Kim, “In Vitro Reconstruction of Brain Tumor Microenvironment,” BioChip Journal 895
2019 13:1, vol. 13, no. 1, pp. 1–7, Mar. 2019, doi: 10.1007/S13206-018-3102-6. 896
[12] J. Kotia, A. Kotwal, and R. Bharti, “Risk Susceptibility of Brain Tumor Classification to Adversarial 897
Attacks,” Advances in Intelligent Systems and Computing, vol. 1061, pp. 181–187, 2020, doi: 898
10.1007/978-3-030-31964-9_17/COVER/. 899
[13] L. Desjardins et al., “Predicting social withdrawal, anxiety and depression symptoms in pediatric 900
brain tumor survivors,” https://doi.org/10.1080/07347332.2018.1535531, vol. 37, no. 1, pp. 22–36, 901
Jan. 2019, doi: 10.1080/07347332.2018.1535531. 902
[14] A. Wadhwa, A. Bhardwaj, and V. Singh Verma, “A review on brain tumor segmentation of MRI 903
images,” Magn Reson Imaging, vol. 61, pp. 247–259, Sep. 2019, doi: 10.1016/J.MRI.2019.05.043. 904
[15] W. Zhang, Y. Wu, B. Yang, S. Hu, L. Wu, and S. Dhelimd, “Overview of Multi-Modal Brain Tumor 905
MR Image Segmentation,” Healthcare 2021, Vol. 9, Page 1051, vol. 9, no. 8, p. 1051, Aug. 2021, 906
doi: 10.3390/HEALTHCARE9081051. 907
[16] P. K. Chahal, S. Pandey, and S. Goel, “A survey on brain tumor detection techniques for MR 908
images,” Multimedia Tools and Applications 2020 79:29, vol. 79, no. 29, pp. 21771–21814, May 909
2020, doi: 10.1007/S11042-020-08898-3. 910
[17] W. L. Bi et al., “Artificial intelligence in cancer imaging: Clinical challenges and applications,” CA 911
Cancer J Clin, vol. 69, no. 2, pp. 127–157, Mar. 2019, doi: 10.3322/CAAC.21552. 912
[18] R. Ranjbarzadeh, A. Bagherian Kasgari, S. Jafarzadeh Ghoushchi, S. Anari, M. Naseri, and M. 913
Bendechache, “Brain tumor segmentation based on deep learning and an attention mechanism using 914
MRI multi-modalities brain images,” Sci Rep, vol. 11, no. 1, p. 10930, Dec. 2021, doi: 915
10.1038/s41598-021-90428-8. 916
[19] M. Afridi, A. Jain, M. Aboian, and S. Payabvash, “Brain Tumor Imaging: Applications of Artificial 917
Intelligence,” Seminars in Ultrasound, CT and MRI, vol. 43, no. 2, pp. 153–169, Apr. 2022, doi: 918
10.1053/J.SULT.2022.02.005. 919
[20] R. Singh, A. Goel, and D. K. Raghuvanshi, “Computer-aided diagnostic network for brain tumor 920
classification employing modulated Gabor filter banks,” The Visual Computer 2020 37:8, vol. 37, 921
no. 8, pp. 2157–2171, Sep. 2020, doi: 10.1007/S00371-020-01977-4. 922
[21] A. Aghamohammadi, R. Ranjbarzadeh, F. Naiemi, M. Mogharrebi, S. Dorosti, and M. Bendechache, 923
“TPCNN: Two-path convolutional neural network for tumor and liver segmentation in CT images 924
using a novel encoding approach,” Expert Syst Appl, vol. 183, p. 115406, Nov. 2021, doi: 925
10.1016/J.ESWA.2021.115406. 926
[22] R. Ranjbarzadeh and S. Baseri Saadi, “Corrigendum to ‘Automated liver and tumor segmentation 927
based on concave and convex points using fuzzy c-means and mean shift clustering’ [Measurement 928
150 (2020) 107086],” Measurement, vol. 151, p. 107230, Feb. 2020, doi: 929
10.1016/J.MEASUREMENT.2019.107230. 930
[23] S. B. Saadi et al., “Osteolysis: A Literature Review of Basic Science and Potential Computer-Based 931
Image Processing Detection Methods,” Comput Intell Neurosci, vol. 2021, 2021, doi: 932
10.1155/2021/4196241. 933
[24] J. Nalepa, M. Marcinkiewicz, and M. Kawulok, “Data Augmentation for Brain-Tumor 934
Segmentation: A Review,” Front Comput Neurosci, vol. 13, p. 83, Dec. 2019, doi: 935
10.3389/FNCOM.2019.00083/BIBTEX. 936
[25] N. J. Shafana and A. Senthilselvi, “Analysis of AI based Brain Tumor Detection and Diagnosis,” 937
Proceedings of the 2021 4th International Conference on Computing and Communications 938
Technologies, ICCCT 2021, pp. 627–633, 2021, doi: 10.1109/ICCCT53315.2021.9711914. 939
[26] S. J. Ghoushchi, R. Ranjbarzadeh, A. H. Dadkhah, Y. Pourasad, and M. Bendechache, “An Extended 940
Approach to Predict Retinopathy in Diabetic Patients Using the Genetic Algorithm and Fuzzy C-941
Means,” Biomed Res Int, vol. 2021, pp. 1–13, Jun. 2021, doi: 10.1155/2021/5597222. 942
[27] A. Naseer, T. Yasir, A. Azhar, T. Shakeel, and K. Zafar, “Computer-Aided Brain Tumor Diagnosis: 943
Performance Evaluation of Deep Learner CNN Using Augmented Brain MRI,” Int J Biomed 944
Imaging, vol. 2021, 2021, doi: 10.1155/2021/5513500. 945
[28] S. Razzaq, M. A. Asghar, and M. J. Khan, “An Improved AI Inspired Brain Tumor Detection,” 3rd 946
International Conference on Communication Technologies, ComTech 2021, pp. 1–6, 2021, doi: 947
10.1109/COMTECH52583.2021.9616648. 948
[29] H. E. M. Abdalla and M. Y. Esmail, “Brain Tumor Detection by using Artificial Neural Network,” 949
2018 International Conference on Computer, Control, Electrical, and Electronics Engineering, 950
ICCCEEE 2018, Oct. 2018, doi: 10.1109/ICCCEEE.2018.8515763. 951
[30] S. Bauer, R. Wiest, L. P. Nolte, and M. Reyes, “A survey of MRI-based medical image analysis for 952
brain tumor studies,” Phys Med Biol, vol. 58, no. 13, p. R97, Jun. 2013, doi: 10.1088/0031-953
9155/58/13/R97. 954
[31] F. Fraioli and S. Punwani, “Clinical and research applications of simultaneous positron emission 955
tomography and MRI,” British Journal of Radiology, vol. 87, no. 1033. British Institute of 956
Radiology, Jan. 2014. doi: 10.1259/bjr.20130464. 957
[32] A. Tiwari, S. Srivastava, and M. Pant, “Brain tumor segmentation and classification from magnetic 958
resonance images: Review of selected methods from 2014 to 2019,” Pattern Recognit Lett, vol. 131, 959
pp. 244–260, Mar. 2020, doi: 10.1016/J.PATREC.2019.11.020. 960
[33] A. Wadhwa, A. Bhardwaj, and V. Singh Verma, “A review on brain tumor segmentation of MRI 961
images,” Magn Reson Imaging, vol. 61, pp. 247–259, Sep. 2019, doi: 10.1016/J.MRI.2019.05.043. 962
[34] S.-G. Choi, C.-B. Sohn, and P. D. Candidate, “Detection of HGG and LGG Brain Tumors using U-963
Net,” Medico Legal Update, vol. 19, no. 1, pp. 560–565, Feb. 2019, doi: 10.37506/MLU.V19I1.978. 964
[35] S. Lapointe, A. Perry, and N. A. Butowski, “Primary brain tumours in adults,” The Lancet, vol. 392, 965
no. 10145, pp. 432–446, Aug. 2018, doi: 10.1016/S0140-6736(18)30990-5. 966
[36] C. J. Valvona, H. L. Fillmore, P. B. Nunn, and G. J. Pilkington, “The Regulation and Function of 967
Lactate Dehydrogenase A: Therapeutic Potential in Brain Tumor,” Brain Pathology, vol. 26, no. 1, 968
pp. 3–17, Jan. 2016, doi: 10.1111/BPA.12299. 969
[37] S.-G. Choi, C.-B. Sohn, and P. D. Candidate, “Detection of HGG and LGG Brain Tumors using U-970
Net,” Medico Legal Update, vol. 19, no. 1, pp. 560–565, Feb. 2019, doi: 10.37506/MLU.V19I1.978. 971
[38] C. I. Hill, C. S. Nixon, J. L. Ruehmeier, and L. M. Wolf, “Brain Tumors,” Phys Ther, vol. 82, no. 972
5, pp. 496–502, May 2002, doi: 10.1093/PTJ/82.5.496. 973
[39] S. Saman and S. Jamjala Narayanan, “Survey on brain tumor segmentation and feature extraction of 974
MR images,” International Journal of Multimedia Information Retrieval 2018 8:2, vol. 8, no. 2, pp. 975
79–99, Dec. 2018, doi: 10.1007/S13735-018-0162-2. 976
[40] S. Bacchi et al., “Deep learning in the detection of high-grade glioma recurrence using multiple MRI 977
sequences: A pilot study,” Journal of Clinical Neuroscience, vol. 70, pp. 11–13, Dec. 2019, doi: 978
10.1016/J.JOCN.2019.10.003. 979
[41] D. A. Hormuth et al., “Translating preclinical MRI methods to clinical oncology,” Journal of 980
Magnetic Resonance Imaging, vol. 50, no. 5, pp. 1377–1392, Nov. 2019, doi: 10.1002/JMRI.26731. 981
[42] J. Amin, M. Sharif, N. Gul, M. Yasmin, and S. A. Shad, “Brain tumor classification based on DWT 982
fusion of MRI sequences using convolutional neural network,” Pattern Recognit Lett, vol. 129, pp. 983
115–122, Jan. 2020, doi: 10.1016/J.PATREC.2019.11.016. 984
[43] S. Saman and S. Jamjala Narayanan, “Survey on brain tumor segmentation and feature extraction of 985
MR images,” International Journal of Multimedia Information Retrieval 2018 8:2, vol. 8, no. 2, pp. 986
79–99, Dec. 2018, doi: 10.1007/S13735-018-0162-2. 987
[44] R. Schettini, “A segmentation algorithm for color images,” Pattern Recognit Lett, vol. 14, no. 6, pp. 988
499–506, Jun. 1993, doi: 10.1016/0167-8655(93)90030-H. 989
[45] T. Kumar, J. Park, M. S. Ali, A. F. M. Shahab Uddin, J. H. Ko, and S.-H. Bae, “Binary-Classifiers-990
Enabled Filters for Semi-Supervised Learning,” IEEE Access, pp. 1–1, 2021, doi: 991
10.1109/ACCESS.2021.3124200. 992
[46] M. Turab, T. Kumar, M. Bendechache, and T. Saber, “Investigating Multi-Feature Selection and 993
Ensembling for Audio Classification,” Jun. 2022, doi: 10.48550/arxiv.2206.07511. 994
[47] S. Bauer, R. Wiest, L. P. Nolte, and M. Reyes, “A survey of MRI-based medical image analysis for 995
brain tumor studies,” Phys Med Biol, vol. 58, no. 13, p. R97, Jun. 2013, doi: 10.1088/0031-996
9155/58/13/R97. 997
[48] A. Baghban, M. Kahani, M. A. Nazari, M. H. Ahmadi, and W. M. Yan, “Sensitivity analysis and 998
application of machine learning methods to predict the heat transfer performance of CNT/water 999
nanofluid flows through coils,” Int J Heat Mass Transf, vol. 128, pp. 825–835, Jan. 2019, doi: 1000
10.1016/J.IJHEATMASSTRANSFER.2018.09.041. 1001
[49] Z. Liu and A. Baghban, “Application of LSSVM for biodiesel production using supercritical ethanol 1002
solvent,” http://dx.doi.org/10.1080/15567036.2017.1380732, vol. 39, no. 17, pp. 1869–1874, Oct. 1003
2017, doi: 10.1080/15567036.2017.1380732. 1004
[50] S. Baseri Saadi, N. Tataei Sarshar, S. Sadeghi, R. Ranjbarzadeh, M. Kooshki Forooshani, and M. 1005
Bendechache, “Investigation of Effectiveness of Shuffled Frog-Leaping Optimizer in Training a 1006
Convolution Neural Network,” J Healthc Eng, vol. 2022, pp. 1–11, Mar. 2022, doi: 1007
10.1155/2022/4703682. 1008
[51] R. Ranjbarzadeh et al., “MRFE-CNN: multi-route feature extraction model for breast tumor 1009
segmentation in Mammograms using a convolutional neural network,” Annals of Operations 1010
Research 2022, pp. 1–22, May 2022, doi: 10.1007/S10479-022-04755-8. 1011
[52] J. Dougherty, R. Kohavi, and M. Sahami, “Supervised and Unsupervised Discretization of 1012
Continuous Features,” Machine Learning Proceedings 1995, pp. 194–202, Jan. 1995, doi: 1013
10.1016/B978-1-55860-377-6.50032-3. 1014
[53] A. Qi et al., “Directional mutation and crossover boosted ant colony optimization with application 1015
to COVID-19 X-ray image segmentation,” Comput Biol Med, vol. 148, p. 105810, Sep. 2022, doi: 1016
10.1016/J.COMPBIOMED.2022.105810. 1017
[54] N. H. Rajini and R. Bhavani, “Classification of MRI brain images using k-nearest neighbor and 1018
artificial neural network,” International Conference on Recent Trends in Information Technology, 1019
ICRTIT 2011, pp. 563–568, 2011, doi: 10.1109/ICRTIT.2011.5972341. 1020
[55] S. Sun and R. Huang, “An adaptive k-nearest neighbor algorithm,” Proceedings - 2010 7th 1021
International Conference on Fuzzy Systems and Knowledge Discovery, FSKD 2010, vol. 1, pp. 91–1022
94, 2010, doi: 10.1109/FSKD.2010.5569740. 1023
[56] M. Havaei, H. Larochelle, P. Poulin, and P. M. Jodoin, “Within-brain classification for brain tumor 1024
segmentation,” International Journal of Computer Assisted Radiology and Surgery 2015 11:5, vol. 1025
11, no. 5, pp. 777–788, Nov. 2015, doi: 10.1007/S11548-015-1311-1. 1026
[57] G. Cinarer and B. G. Emiroglu, “Classificatin of Brain Tumors by Machine Learning Algorithms,” 1027
3rd International Symposium on Multidisciplinary Studies and Innovative Technologies, ISMSIT 1028
2019 - Proceedings, Oct. 2019, doi: 10.1109/ISMSIT.2019.8932878. 1029
[58] D. M. Kumar, D. Satyanarayana, and M. N. G. Prasad, “MRI brain tumor detection using optimal 1030
possibilistic fuzzy C-means clustering algorithm and adaptive k-nearest neighbor classifier,” 1031
Journal of Ambient Intelligence and Humanized Computing 2020 12:2, vol. 12, no. 2, pp. 2867–1032
2880, Sep. 2020, doi: 10.1007/S12652-020-02444-7. 1033
[59] A. S. Remya Ajai and S. Gopalan, “Analysis of active contours without edge-based segmentation 1034
technique for brain tumor classification using svm and knn classifiers,” Lecture Notes in Electrical 1035
Engineering, vol. 656, pp. 1–10, 2020, doi: 10.1007/978-981-15-3992-3_1/COVER/. 1036
[60] R. H. Ramdlon, E. Martiana Kusumaningtyas, and T. Karlita, “Brain Tumor Classification Using 1037
MRI Images with K-Nearest Neighbor Method,” IES 2019 - International Electronics Symposium: 1038
The Role of Techno-Intelligence in Creating an Open Energy System Towards Energy Democracy, 1039
Proceedings, pp. 660–667, Sep. 2019, doi: 10.1109/ELECSYM.2019.8901560. 1040
[61] V. V. P. Wibowo, Z. Rustam, and J. Pandelaki, “Classification of Brain Tumor Using K-Nearest 1041
Neighbor-Genetic Algorithm and Support Vector Machine-Genetic Algorithm Methods,” 2021 1042
International Conference on Decision Aid Sciences and Application, DASA 2021, pp. 1077–1081, 1043
2021, doi: 10.1109/DASA53625.2021.9682341. 1044
[62] A. Shmilovici, “Support Vector Machines,” Data Mining and Knowledge Discovery Handbook, pp. 1045
231–247, 2009, doi: 10.1007/978-0-387-09823-4_12. 1046
[63] H. B. Nandpuru, S. S. Salankar, and V. R. Bora, “MRI brain cancer classification using support 1047
vector machine,” 2014 IEEE Students’ Conference on Electrical, Electronics and Computer 1048
Science, SCEECS 2014, 2014, doi: 10.1109/SCEECS.2014.6804439. 1049
[64] J. Amin, M. Sharif, M. Yasmin, and S. L. Fernandes, “A distinctive approach in brain tumor 1050
detection and classification using MRI,” Pattern Recognit Lett, vol. 139, pp. 118–127, Nov. 2020, 1051
doi: 10.1016/J.PATREC.2017.10.036. 1052
[65] M. Padlia and J. Sharma, “Fractional sobel filter based brain tumor detection and segmentation using 1053
statistical features and SVM,” Lecture Notes in Electrical Engineering, vol. 511, pp. 161–175, 2019, 1054
doi: 10.1007/978-981-13-0776-8_15/COVER/. 1055
[66] M. O. Khairandish, M. Sharma, V. Jain, J. M. Chatterjee, and N. Z. Jhanjhi, “A Hybrid CNN-SVM 1056
Threshold Segmentation Approach for Tumor Detection and Classification of MRI Brain Images,” 1057
IRBM, Jun. 2021, doi: 10.1016/J.IRBM.2021.06.003. 1058
[67] C. S. Rao and K. Karunakara, “Efficient Detection and Classification of Brain Tumor using Kernel 1059
based SVM for MRI,” Multimedia Tools and Applications 2022 81:5, vol. 81, no. 5, pp. 7393–7417, 1060
Jan. 2022, doi: 10.1007/S11042-021-11821-Z. 1061
[68] M. H. O. Rashid, M. A. Mamun, M. A. Hossain, and M. P. Uddin, “Brain Tumor Detection Using 1062
Anisotropic Filtering, SVM Classifier and Morphological Operation from MR Images,” 1063
International Conference on Computer, Communication, Chemical, Material and Electronic 1064
Engineering, IC4ME2 2018, Sep. 2018, doi: 10.1109/IC4ME2.2018.8465613. 1065
[69] S. Deepak and P. M. Ameer, “Automated Categorization of Brain Tumor from MRI Using CNN 1066
features and SVM,” Journal of Ambient Intelligence and Humanized Computing 2020 12:8, vol. 12, 1067
no. 8, pp. 8357–8369, Oct. 2020, doi: 10.1007/S12652-020-02568-W. 1068
[70] Y. Qi, “Random Forest for Bioinformatics,” Ensemble Machine Learning, pp. 307–323, 2012, doi: 1069
10.1007/978-1-4419-9326-7_11. 1070
[71] T. M. Oshiro, P. S. Perez, and J. A. Baranauskas, “How many trees in a random forest?,” Lecture 1071
Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture 1072
Notes in Bioinformatics), vol. 7376 LNAI, pp. 154–168, 2012, doi: 10.1007/978-3-642-31537-1073
4_13/COVER. 1074
[72] L. Lefkovits, S. Lefkovits, and L. Szilágyi, “Brain Tumor Segmentation with Optimized Random 1075
Forest,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial 1076
Intelligence and Lecture Notes in Bioinformatics), vol. 10154 LNCS, pp. 88–99, Oct. 2016, doi: 1077
10.1007/978-3-319-55524-9_9. 1078
[73] A. Ellwaa et al., “Brain tumor segmantation using random forest trained on iteratively selected 1079
patients,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial 1080
Intelligence and Lecture Notes in Bioinformatics), vol. 10154 LNCS, pp. 129–137, 2016, doi: 1081
10.1007/978-3-319-55524-9_13/COVER/. 1082
[74] R. Anitha and D. Siva Sundhara Raja, “Development of computer-aided approach for brain tumor 1083
detection using random forest classifier,” Int J Imaging Syst Technol, vol. 28, no. 1, pp. 48–53, Mar. 1084
2018, doi: 10.1002/IMA.22255. 1085
[75] T. Yang, J. Song, and L. Li, “A deep learning model integrating SK-TPCNN and random forests for 1086
brain tumor segmentation in MRI,” Biocybern Biomed Eng, vol. 39, no. 3, pp. 613–623, Jul. 2019, 1087
doi: 10.1016/J.BBE.2019.06.003. 1088
[76] R. Rajagopal, “Glioma brain tumor detection and segmentation using weighting random forest 1089
classifier with optimized ant colony features,” Int J Imaging Syst Technol, vol. 29, no. 3, pp. 353–1090
359, Sep. 2019, doi: 10.1002/IMA.22331. 1091
[77] A. J. Myles, R. N. Feudale, Y. Liu, N. A. Woody, and S. D. Brown, “An introduction to decision 1092
tree modeling,” J Chemom, vol. 18, no. 6, pp. 275–285, Jun. 2004, doi: 10.1002/CEM.873. 1093
[78] W. Wang, A. Tu, and F. Bergholm, “Improved Minimum Spanning Tree based Image Segmentation 1094
with Guided Matting,” KSII Transactions on Internet and Information Systems (TIIS), vol. 16, no. 1095
1, pp. 211–230, Jan. 2022, doi: 10.3837/TIIS.2022.01.012. 1096
[79] Y. Y. Song and Y. Lu, “Decision tree methods: applications for classification and prediction,” 1097
Shanghai Arch Psychiatry, vol. 27, no. 2, p. 130, Apr. 2015, doi: 10.11919/J.ISSN.1002-1098
0829.215044. 1099
[80] “Tumor Detection and Classification using Decision Tree in Brain MRI | Semantic Scholar.” 1100
[81] A. Chaddad, P. O. Zinn, and R. R. Colen, “Brain tumor identification using Gaussian Mixture Model 1101
features and Decision Trees classifier,” 2014 48th Annual Conference on Information Sciences and 1102
Systems, CISS 2014, 2014, doi: 10.1109/CISS.2014.6814077. 1103
[82] L. Hussain, S. Saeed, I. A. Awan, A. Idris, M. S. A. Nadeem, and Q.-A. Chaudhry, “Detecting Brain 1104
Tumor using Machines Learning Techniques Based on Different Features Extracting Strategies,” 1105
Current Medical Imaging Formerly Current Medical Imaging Reviews, vol. 15, no. 6, pp. 595–606, 1106
Jul. 2019, doi: 10.2174/1573405614666180718123533. 1107
[83] M. Thayumanavan and A. Ramasamy, “An efficient approach for brain tumor detection and 1108
segmentation in MR brain images using random forest classifier:,” 1109
https://doi.org/10.1177/1063293X211010542, vol. 29, no. 3, pp. 266–274, Apr. 2021, doi: 1110
10.1177/1063293X211010542. 1111
[84] P. Rajendran and M. Madheswaran, “Hybrid Medical Image Classification Using Association Rule 1112
Mining with Decision Tree Algorithm,” vol. 2, Jan. 2010, doi: 10.48550/arxiv.1001.3503. 1113
[85] Z. Zhang, “Artificial Neural Network,” Multivariate Time Series Analysis in Climate and 1114
Environmental Research, pp. 1–35, 2018, doi: 10.1007/978-3-319-67340-0_1. 1115
[86] S. Tang and F. Yu, “Construction and verification of retinal vessel segmentation algorithm for color 1116
fundus image under BP neural network model,” Journal of Supercomputing, vol. 77, no. 4, pp. 1117
3870–3884, Apr. 2021, doi: 10.1007/S11227-020-03422-8/FIGURES/4. 1118
[87] B. He et al., “Image segmentation algorithm of lung cancer based on neural network model,” Expert 1119
Syst, vol. 39, no. 3, p. e12822, Mar. 2022, doi: 10.1111/EXSY.12822. 1120
[88] M. Mishra and M. Srivastava, “A view of Artificial Neural Network,” 2014 International 1121
Conference on Advances in Engineering and Technology Research, ICAETR 2014, 2014, doi: 1122
10.1109/ICAETR.2014.7012785. 1123
[89] T. Chithambaram and K. Perumal, “Brain tumor segmentation using genetic algorithm and ANN 1124
techniques,” IEEE International Conference on Power, Control, Signals and Instrumentation 1125
Engineering, ICPCSI 2017, pp. 970–982, Jun. 2018, doi: 10.1109/ICPCSI.2017.8391855. 1126
[90] Virupakshappa and B. Amarapur, “Computer-aided diagnosis applied to MRI images of brain tumor 1127
using cognition based modified level set and optimized ANN classifier,” Multimedia Tools and 1128
Applications 2018 79:5, vol. 79, no. 5, pp. 3571–3599, Jun. 2018, doi: 10.1007/S11042-018-6176-1129
1. 1130
[91] Virupakshappa and B. Amarapur, “An Automated Approach for Brain Tumor Identification using 1131
ANN Classifier,” International Conference on Current Trends in Computer, Electrical, Electronics 1132
and Communication, CTCEEC 2017, pp. 1011–1016, Sep. 2018, doi: 1133
10.1109/CTCEEC.2017.8455154. 1134
[92] G. I. Webb, “Naïve Bayes,” Encyclopedia of Machine Learning and Data Mining, pp. 1–2, 2016, 1135
doi: 10.1007/978-1-4899-7502-7_581-1. 1136
[93] W. Li, G. G. Wang, and A. H. Gandomi, “A Survey of Learning-Based Intelligent Optimization 1137
Algorithms,” Archives of Computational Methods in Engineering 2021 28:5, vol. 28, no. 5, pp. 1138
3781–3799, Feb. 2021, doi: 10.1007/S11831-021-09562-1. 1139
[94] L. Jiang, H. Zhang, and Z. Cai, “A novel bayes model: Hidden naive bayes,” IEEE Trans Knowl 1140
Data Eng, vol. 21, no. 10, pp. 1361–1371, Oct. 2009, doi: 10.1109/TKDE.2008.234. 1141
[95] G. Kaur and A. Oberoi, “Novel Approach for Brain Tumor Detection Based on Naïve Bayes 1142
Classification,” Advances in Intelligent Systems and Computing, vol. 1042, pp. 451–462, 2020, doi: 1143
10.1007/978-981-32-9949-8_31/COVER/. 1144
[96] A. R. Raju, P. Suresh, and R. R. Rao, “Bayesian HCS-based multi-SVNN: A classification approach 1145
for brain tumor segmentation and classification using Bayesian fuzzy clustering,” Biocybern Biomed 1146
Eng, vol. 38, no. 3, pp. 646–660, Jan. 2018, doi: 10.1016/J.BBE.2018.05.001. 1147
[97] E. E. Ulku and A. Y. Camurcu, “Computer aided brain tumor detection with histogram equalization 1148
and morphological image processing techniques,” 2013 International Conference on Electronics, 1149
Computer and Computation, ICECCO 2013, pp. 48–51, 2013, doi: 1150
10.1109/ICECCO.2013.6718225. 1151
[98] T. Kohonen, “Learning Vector Quantization,” pp. 245–261, 2001, doi: 10.1007/978-3-642-56927-1152
2_6. 1153
[99] D. Nova and P. A. Estévez, “A review of learning vector quantization classifiers,” Neural 1154
Computing and Applications 2013 25:3, vol. 25, no. 3, pp. 511–524, Dec. 2013, doi: 1155
10.1007/S00521-013-1535-3. 1156
[100] T. Liu, C. Chen, X. Shi, and C. Liu, “Evaluation of Raman spectra of human brain tumor tissue 1157
using the learning vector quantization neural network,” Laser Phys, vol. 26, no. 5, p. 055606, Apr. 1158
2016, doi: 10.1088/1054-660X/26/5/055606. 1159
[101] R. Sonavane, P. Sonar, and S. Sutar, “Classification of MRI brain tumor and mammogram images 1160
using learning vector quantization neural network,” Proceedings of 2017 3rd IEEE International 1161
Conference on Sensing, Signal Processing and Security, ICSSS 2017, pp. 301–307, Oct. 2017, doi: 1162
10.1109/SSPS.2017.8071610. 1163
[102] R. S. Sonavane and A. S. Sonavane, “Classification of MRI brain tumor and mammogram images 1164
using adaboost and learning vector quantization neural network,” 2020 International Conference on 1165
Convergence to Digital World - Quo Vadis, ICCDW 2020, Feb. 2020, doi: 1166
10.1109/ICCDW45521.2020.9318645. 1167
[103] T. Hastie, R. Tibshirani, and J. Friedman, “Unsupervised Learning,” Springer, New York, NY, 2009, 1168
pp. 485–585. doi: 10.1007/978-0-387-84858-7_14. 1169
[104] H. Su et al., “Multilevel threshold image segmentation for COVID-19 chest radiography: A 1170
framework using horizontal and vertical multiverse optimization,” Comput Biol Med, vol. 146, p. 1171
105618, Jul. 2022, doi: 10.1016/J.COMPBIOMED.2022.105618. 1172
[105] S. Anari, N. Tataei Sarshar, N. Mahjoori, S. Dorosti, and A. Rezaie, “Review of Deep Learning 1173
Approaches for Thyroid Cancer Diagnosis,” Math Probl Eng, vol. 2022, pp. 1–8, Aug. 2022, doi: 1174
10.1155/2022/5052435. 1175
[106] D. Gao, G. G. Wang, and W. Pedrycz, “Solving Fuzzy Job-Shop Scheduling Problem Using de 1176
Algorithm Improved by a Selection Mechanism,” IEEE Transactions on Fuzzy Systems, vol. 28, no. 1177
12, pp. 3265–3275, Dec. 2020, doi: 10.1109/TFUZZ.2020.3003506. 1178
[107] G. G. Wang, D. Gao, and W. Pedrycz, “Solving Multi-Objective Fuzzy Job-shop Scheduling 1179
Problem by a Hybrid Adaptive Differential Evolution Algorithm,” IEEE Trans Industr Inform, Dec. 1180
2022, doi: 10.1109/TII.2022.3165636. 1181
[108] M. A. T. Figueiredo and A. K. Jain, “Unsupervised learning of finite mixture models,” IEEE Trans 1182
Pattern Anal Mach Intell, vol. 24, no. 3, pp. 381–396, Mar. 2002, doi: 10.1109/34.990138. 1183
[109] K. Krishna and M. N. Murty, “Genetic K-means algorithm,” IEEE Transactions on Systems, Man, 1184
and Cybernetics, Part B: Cybernetics, vol. 29, no. 3, pp. 433–439, 1999, doi: 10.1109/3477.764879. 1185
[110] A. Likas, N. Vlassis, and J. J. Verbeek, “The global k-means clustering algorithm,” Pattern 1186
Recognit, vol. 36, no. 2, pp. 451–461, Feb. 2003, doi: 10.1016/S0031-3203(02)00060-2. 1187
[111] R. Khilkhal and M. Ismael, “Brain Tumor Segmentation Utilizing Thresholding and K-Means 1188
Clustering,” pp. 43–48, Jun. 2022, doi: 10.1109/MICEST54286.2022.9790103. 1189
[112] M. K. Islam, M. S. Ali, M. S. Miah, M. M. Rahman, M. S. Alam, and M. A. Hossain, “Brain tumor 1190
detection in MR image using superpixels, principal component analysis and template based K-1191
means clustering algorithm,” Machine Learning with Applications, vol. 5, p. 100044, Sep. 2021, 1192
doi: 10.1016/J.MLWA.2021.100044. 1193
[113] D. M. Kumar, D. Satyanarayana, and M. N. G. Prasad, “An improved Gabor wavelet transform and 1194
rough K-means clustering algorithm for MRI brain tumor image segmentation,” Multimedia Tools 1195
and Applications 2020 80:5, vol. 80, no. 5, pp. 6939–6957, Oct. 2020, doi: 10.1007/S11042-020-1196
09635-6. 1197
[114] T. C. Havens, J. C. Bezdek, C. Leckie, L. O. Hall, and M. Palaniswami, “Fuzzy c-Means algorithms 1198
for very large data,” IEEE Transactions on Fuzzy Systems, vol. 20, no. 6, pp. 1130–1146, 2012, doi: 1199
10.1109/TFUZZ.2012.2201485. 1200
[115] S. Askari, “Fuzzy C-Means clustering algorithm for data with unequal cluster sizes and 1201
contaminated with noise and outliers: Review and development,” Expert Syst Appl, vol. 165, p. 1202
113856, Mar. 2021, doi: 10.1016/J.ESWA.2020.113856. 1203
[116] H. Izakian and A. Abraham, “Fuzzy C-means and fuzzy swarm for fuzzy clustering problem,” 1204
Expert Syst Appl, vol. 38, no. 3, pp. 1835–1838, Mar. 2011, doi: 10.1016/J.ESWA.2010.07.112. 1205
[117] R. Sindhiya Devi, B. Perumal, and M. Pallikonda Rajasekaran, “A hybrid deep learning based brain 1206
tumor classification and segmentation by stationary wavelet packet transform and adaptive kernel 1207
fuzzy c means clustering,” Advances in Engineering Software, vol. 170, p. 103146, Aug. 2022, doi: 1208
10.1016/J.ADVENGSOFT.2022.103146. 1209
[118] S. Debnath, F. A. Talukdar, and M. Islam, “Combination of contrast enhanced fuzzy c-means 1210
(CEFCM) clustering and pixel based voxel mapping technique (PBVMT) for three dimensional 1211
brain tumour detection,” Journal of Ambient Intelligence and Humanized Computing 2020 12:2, 1212
vol. 12, no. 2, pp. 2421–2433, Aug. 2020, doi: 10.1007/S12652-020-02366-4. 1213
[119] C. J. J. Sheela and G. Suganthi, “Accurate MRI brain tumor segmentation based on rotating 1214
triangular section with fuzzy C- means optimization,” Sādhanā 2021 46:4, vol. 46, no. 4, pp. 1–20, 1215
Oct. 2021, doi: 10.1007/S12046-021-01744-8. 1216
[120] M. Soleymanifard and M. Hamghalam, “Multi-stage glioma segmentation for tumour grade 1217
classification based on multiscale fuzzy C-means,” Multimedia Tools and Applications 2022 81:6, 1218
vol. 81, no. 6, pp. 8451–8470, Feb. 2022, doi: 10.1007/S11042-022-12326-Z. 1219
[121] X. T. Yuan, B. G. Hu, and R. He, “Agglomerative mean-shift clustering,” IEEE Trans Knowl Data 1220
Eng, vol. 24, no. 2, pp. 209–219, 2012, doi: 10.1109/TKDE.2010.232. 1221
[122] R. Ranjbarzadeh and S. B. Saadi, “Automated liver and tumor segmentation based on concave and 1222
convex points using fuzzy c-means and mean shift clustering,” Measurement, vol. 150, p. 107086, 1223
Jan. 2020, doi: 10.1016/J.MEASUREMENT.2019.107086. 1224
[123] K. L. Wu and M. S. Yang, “Mean shift-based clustering,” Pattern Recognit, vol. 40, no. 11, pp. 1225
3035–3052, Nov. 2007, doi: 10.1016/J.PATCOG.2007.02.006. 1226
[124] R. B. Vallabhaneni and V. Rajesh, “Brain tumour detection using mean shift clustering and GLCM 1227
features with edge adaptive total variation denoising technique,” Alexandria Engineering Journal, 1228
vol. 57, no. 4, pp. 2387–2392, Dec. 2018, doi: 10.1016/J.AEJ.2017.09.011. 1229
[125] B. Singh and P. Aggarwal, “Detection of brain tumor using modified mean-shift based fuzzy c-mean 1230
segmentation from MRI Images,” 2017 8th IEEE Annual Information Technology, Electronics and 1231
Mobile Communication Conference, IEMCON 2017, pp. 536–545, Nov. 2017, doi: 1232
10.1109/IEMCON.2017.8117123. 1233
[126] J. H. Kim, S. Lee, G. S. Lee, Y. S. Park, and Y. P. Hong, “Using a Method Based on a Modified K-1234
Means Clustering and Mean Shift Segmentation to Reduce File Sizes and Detect Brain Tumors from 1235
Magnetic Resonance (MRI) Images,” Wireless Personal Communications 2016 89:3, vol. 89, no. 3, 1236
pp. 993–1008, Jun. 2016, doi: 10.1007/S11277-016-3420-8. 1237
[127] F. Murtagh and P. Contreras, “Algorithms for hierarchical clustering: an overview,” Wiley 1238
Interdiscip Rev Data Min Knowl Discov, vol. 2, no. 1, pp. 86–97, Jan. 2012, doi: 10.1002/WIDM.53. 1239
[128] F. Nielsen, “Hierarchical Clustering,” pp. 195–211, 2016, doi: 10.1007/978-3-319-21903-5_8. 1240
[129] T. Hiratsuka et al., “Hierarchical Cluster and Region of Interest Analyses Based on Mass 1241
Spectrometry Imaging of Human Brain Tumours,” Scientific Reports 2020 10:1, vol. 10, no. 1, pp. 1242
1–11, Apr. 2020, doi: 10.1038/s41598-020-62176-8. 1243
[130] G. Tamilmani and S. Sivakumari, “Early detection of brain cancer using association allotment 1244
hierarchical clustering,” Int J Imaging Syst Technol, vol. 29, no. 4, pp. 617–632, Dec. 2019, doi: 1245
10.1002/IMA.22346. 1246
[131] K. Khan, S. U. Rehman, K. Aziz, S. Fong, S. Sarasvady, and A. Vishwa, “DBSCAN: Past, present 1247
and future,” 5th International Conference on the Applications of Digital Information and Web 1248
Technologies, ICADIWT 2014, pp. 232–238, 2014, doi: 10.1109/ICADIWT.2014.6814687. 1249
[132] E. Schubert, J. Sander, M. Ester, H. P. Kriegel, and X. Xu, “DBSCAN Revisited, Revisited,” ACM 1250
Transactions on Database Systems (TODS), vol. 42, no. 3, Jul. 2017, doi: 10.1145/3068335. 1251
[133] “ANALYSIS OF BRAIN TUMOR CLASSIFICATION BY USING MULTIPLE CLUSTERING 1252
ALGORITHMS | Semantic Scholar.” 1253
[134] “[PDF] Segmentation of Brain Tumour from MRI image – Analysis of K-means and DBSCAN 1254
Clustering | Semantic Scholar.” 1255
[135] M. A. T. Figueiredo and A. K. Jain, “Unsupervised learning of finite mixture models,” IEEE Trans 1256
Pattern Anal Mach Intell, vol. 24, no. 3, pp. 381–396, Mar. 2002, doi: 10.1109/34.990138. 1257
[136] A. Chaddad, “Automated feature extraction in brain tumor by magnetic resonance imaging using 1258
Gaussian mixture models,” Journal of Biomedical Imaging, vol. 2015, Jan. 2015, doi: 1259
10.1155/2015/868031. 1260
[137] A. A. Pravitasari et al., “MRI-based brain tumor segmentation using Gaussian mixture model with 1261
reversible jump Markov chain Monte Carlo algorithm,” AIP Conf Proc, vol. 2194, no. 1, p. 020085, 1262
Dec. 2019, doi: 10.1063/1.5139817. 1263
[138] S. Bonte, I. Goethals, and R. Van Holen, “Machine learning based brain tumour segmentation on 1264
limited data using local texture and abnormality,” Comput Biol Med, vol. 98, pp. 39–47, Jul. 2018, 1265
doi: 10.1016/J.COMPBIOMED.2018.05.005. 1266
[139] N. B. Bahadure, A. K. Ray, and H. P. Thethi, “Comparative Approach of MRI-Based Brain Tumor 1267
Segmentation and Classification Using Genetic Algorithm,” Journal of Digital Imaging 2018 31:4, 1268
vol. 31, no. 4, pp. 477–489, Jan. 2018, doi: 10.1007/S10278-018-0050-6. 1269
[140] A. Srinivasa Reddy and P. Chenna Reddy, “MRI brain tumor segmentation and prediction using 1270
modified region growing and adaptive SVM,” Soft Computing 2021 25:5, vol. 25, no. 5, pp. 4135–1271
4148, Jan. 2021, doi: 10.1007/S00500-020-05493-4. 1272
[141] X. Xie, “A K-Nearest Neighbor Technique for Brain Tumor Segmentation Using Minkowski 1273
Distance,” J Med Imaging Health Inform, vol. 8, no. 2, pp. 180–185, Jan. 2018, doi: 1274
10.1166/JMIHI.2018.2285. 1275
[142] P. M. Siva Raja and A. V. rani, “Brain tumor classification using a hybrid deep autoencoder with 1276
Bayesian fuzzy clustering-based segmentation approach,” Biocybern Biomed Eng, vol. 40, no. 1, 1277
pp. 440–453, Jan. 2020, doi: 10.1016/J.BBE.2020.01.006. 1278
[143] U. Ilhan and A. Ilhan, “Brain tumor segmentation based on a new threshold approach,” Procedia 1279
Comput Sci, vol. 120, pp. 580–587, Jan. 2017, doi: 10.1016/J.PROCS.2017.11.282. 1280
[144] A. Aslam, E. Khan, and M. M. S. Beg, “Improved Edge Detection Algorithm for Brain Tumor 1281
Segmentation,” Procedia Comput Sci, vol. 58, pp. 430–437, Jan. 2015, doi: 1282
10.1016/J.PROCS.2015.08.057. 1283
[145] A. Kermi, K. Andjouh, and F. Zidane, “Fully automated brain tumour segmentation system in 3D-1284
MRI using symmetry analysis of brain and level sets,” IET Image Process, vol. 12, no. 11, pp. 1964–1285
1971, Nov. 2018, doi: 10.1049/IET-IPR.2017.1124. 1286
[146] C. J. J. Sheela and G. Suganthi, “Morphological edge detection and brain tumor segmentation in 1287
Magnetic Resonance (MR) images based on region growing and performance evaluation of 1288
modified Fuzzy C-Means (FCM) algorithm,” Multimedia Tools and Applications 2020 79:25, vol. 1289
79, no. 25, pp. 17483–17496, Feb. 2020, doi: 10.1007/S11042-020-08636-9. 1290
[147] A. R. Khan, S. Khan, M. Harouni, R. Abbasi, S. Iqbal, and Z. Mehmood, “Brain tumor segmentation 1291
using K-means clustering and deep learning with synthetic data augmentation for classification,” 1292
Microsc Res Tech, vol. 84, no. 7, pp. 1389–1399, Jul. 2021, doi: 10.1002/JEMT.23694. 1293
[148] F. ŞİŞİK and E. SERT, “Brain tumor segmentation approach based on the extreme learning machine 1294
and significantly fast and robust fuzzy C-means clustering algorithms running on Raspberry Pi 1295
hardware,” Med Hypotheses, vol. 136, p. 109507, Mar. 2020, doi: 10.1016/J.MEHY.2019.109507. 1296
[149] A. Khosravanian, M. Rahmanimanesh, P. Keshavarzi, and S. Mozaffari, “Fast Level set Method for 1297
Glioma Brain Tumor Segmentation based on Superpixel Fuzzy Clustering and Lattice Boltzmann 1298
Method,” Comput Methods Programs Biomed, vol. 198, p. 105809, Oct. 2020, doi: 1299
10.1016/j.cmpb.2020.105809. 1300
[150] K. Suzuki, “Overview of deep learning in medical imaging,” Radiological Physics and Technology 1301
2017 10:3, vol. 10, no. 3, pp. 257–273, Jul. 2017, doi: 10.1007/S12194-017-0406-5. 1302
[151] M. T. S. T. R. B. and M. B. Teerath Kumar et al., “Forged Character Detection Datasets: Passports, 1303
Driving Licences and Visa Stickers,” InternationalJournal of ArtificialIntelligence 1304
Applications(IJAIA), Vol.13,No.2, vol. 13, no. 2, p. 21, Mar. 2022, doi: 10.5121/IJAIA.2022.13202. 1305
[152] A. Aiman, Y. Shen, M. Bendechache, I. Inayat, and T. Kumar, “AUDD: Audio Urdu Digits Dataset 1306
for Automatic Audio Urdu Digit Recognition,” Applied Sciences 2021, Vol. 11, Page 8842, vol. 11, 1307
no. 19, p. 8842, Sep. 2021, doi: 10.3390/APP11198842. 1308
[153] 박진배*, T. Kumar, 경희대학교배성호, J. Park, and S.-H. Bae, “Search of an Optimal Sound 1309
Augmentation Policy for Environmental Sound Classification with Deep Neural Networks,” 1310
Proceedings of the Korean Society of Broadcast Engineers Conference, pp. 18–21, 2020. 1311
[154] S. Aleem, T. Kumar, S. Little, M. Bendechache, R. Brennan, and K. McGuinness, “Random Data 1312
Augmentation based Enhancement: AGeneralized Enhancement Approach for Medical Datasets.” 1313
Dec. 17, 2021. 1314
[155] M. I. Razzak, S. Naz, and A. Zaib, “Deep Learning for Medical Image Processing: Overview, 1315
Challenges and the Future BT - Classification in BioApps: Automation of Decision Making,” 1316
Springer, vol. 26, pp. 323–350, 2018. 1317
[156] R. Ranjbarzadeh et al., “Nerve optic segmentation in CT images using a deep learning model and a 1318
texture descriptor,” Complex & Intelligent Systems 2022, pp. 1–15, Feb. 2022, doi: 10.1007/S40747-1319
022-00694-W. 1320
[157] S. Wang et al., “Multi-Scale Context-Guided Deep Network for Automated Lesion Segmentation 1321
with Endoscopy Images of Gastrointestinal Tract,” IEEE J Biomed Health Inform, vol. 25, no. 2, 1322
pp. 514–525, Feb. 2021, doi: 10.1109/JBHI.2020.2997760. 1323
[158] S. Albawi, T. A. Mohammed, and S. Al-Zawi, “Understanding of a convolutional neural network,” 1324
in Proceedings of 2017 International Conference on Engineering and Technology, ICET 2017, Mar. 1325
2018, vol. 2018-Janua, pp. 1–6. doi: 10.1109/ICEngTechnol.2017.8308186. 1326
[159] R. Yamashita, M. Nishio, R. K. G. Do, and K. Togashi, “Convolutional neural networks: an 1327
overview and application in radiology,” Insights into Imaging, vol. 9, no. 4. Springer Verlag, pp. 1328
611–629, Aug. 2018. doi: 10.1007/s13244-018-0639-9. 1329
[160] I. A. El Kader, G. Xu, Z. Shuai, S. Saminu, I. Javaid, and I. S. Ahmad, “Differential Deep 1330
Convolutional Neural Network Model for Brain Tumor Classification,” Brain Sciences 2021, Vol. 1331
11, Page 352, vol. 11, no. 3, p. 352, Mar. 2021, doi: 10.3390/BRAINSCI11030352. 1332
[161] N. Bacanin, T. Bezdan, K. Venkatachalam, and F. Al-Turjman, “Optimized convolutional neural 1333
network by firefly algorithm for magnetic resonance image classification of glioma brain tumor 1334
grade,” Journal of Real-Time Image Processing 2021 18:4, vol. 18, no. 4, pp. 1085–1098, Apr. 1335
2021, doi: 10.1007/S11554-021-01106-X. 1336
[162] J. Wang et al., “DFP-ResUNet:Convolutional Neural Network with a Dilated Convolutional Feature 1337
Pyramid for Multimodal Brain Tumor Segmentation,” Comput Methods Programs Biomed, vol. 1338
208, p. 106208, Sep. 2021, doi: 10.1016/J.CMPB.2021.106208. 1339
[163] A. Gurunathan and B. Krishnan, “Detection and diagnosis of brain tumors using deep learning 1340
convolutional neural networks,” Int J Imaging Syst Technol, vol. 31, no. 3, pp. 1174–1184, Sep. 1341
2021, doi: 10.1002/IMA.22532. 1342
[164] T. Mikolov, S. Kombrink, L. Burget, J. Černocký, and S. Khudanpur, “Extensions of recurrent 1343
neural network language model,” ICASSP, IEEE International Conference on Acoustics, Speech 1344
and Signal Processing - Proceedings, pp. 5528–5531, 2011, doi: 10.1109/ICASSP.2011.5947611. 1345
[165] M. Lukoševičius and H. Jaeger, “Reservoir computing approaches to recurrent neural network 1346
training,” Comput Sci Rev, vol. 3, no. 3, pp. 127–149, Aug. 2009, doi: 1347
10.1016/J.COSREV.2009.03.005. 1348
[166] A. C. Tsoi, “Recurrent neural network architectures: An overview,” pp. 1–26, 1998, doi: 1349
10.1007/BFB0053993. 1350
[167] J. G. SivaSai, P. N. Srinivasu, M. N. Sindhuri, K. Rohitha, and S. Deepika, “An automated 1351
segmentation of brain MR image through fuzzy recurrent neural network,” Studies in Computational 1352
Intelligence, vol. 903, pp. 163–179, 2021, doi: 10.1007/978-981-15-5495-7_9/COVER/. 1353
[168] Y. Zhou et al., “Holistic brain tumor screening and classification based on densenet and recurrent 1354
neural network,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial 1355
Intelligence and Lecture Notes in Bioinformatics), vol. 11383 LNCS, pp. 208–217, 2019, doi: 1356
10.1007/978-3-030-11723-8_21/COVER/. 1357
[169] S. S. Begum and D. R. Lakshmi, “Combining optimal wavelet statistical texture and recurrent neural 1358
network for tumour detection and classification over MRI,” Multimedia Tools and Applications 1359
2020 79:19, vol. 79, no. 19, pp. 14009–14030, Feb. 2020, doi: 10.1007/S11042-020-08643-W. 1360
[170] G. Van Houdt, C. Mosquera, and G. Nápoles, “A review on the long short-term memory model,” 1361
Artificial Intelligence Review 2020 53:8, vol. 53, no. 8, pp. 5929–5955, May 2020, doi: 1362
10.1007/S10462-020-09838-1. 1363
[171] A. Graves, “Long Short-Term Memory,” pp. 37–45, 2012, doi: 10.1007/978-3-642-24797-2_4. 1364
[172] E. Dandıl and S. Karaca, “Detection of pseudo brain tumors via stacked LSTM neural networks 1365
using MR spectroscopy signals,” Biocybern Biomed Eng, vol. 41, no. 1, pp. 173–195, Jan. 2021, 1366
doi: 10.1016/J.BBE.2020.12.003. 1367
[173] F. Xu, H. Ma, J. Sun, R. Wu, X. Liu, and Y. Kong, “LSTM Multi-modal UNet for Brain Tumor 1368
Segmentation,” 2019 IEEE 4th International Conference on Image, Vision and Computing, ICIVC 1369
2019, pp. 236–240, Jul. 2019, doi: 10.1109/ICIVC47709.2019.8981027. 1370
[174] I. Shahzadi, F. Meriadeau, T. B. Tang, and A. Quyyum, “CNN-LSTM: Cascaded framework for 1371
brain tumour classification,” 2018 IEEE EMBS Conference on Biomedical Engineering and 1372
Sciences, IECBES 2018 - Proceedings, pp. 633–637, Jan. 2019, doi: 1373
10.1109/IECBES.2018.8626704. 1374
[175] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and A. A. Bharath, “Generative 1375
Adversarial Networks: An Overview,” IEEE Signal Process Mag, vol. 35, no. 1, pp. 53–65, Jan. 1376
2018, doi: 10.1109/MSP.2017.2765202. 1377
[176] K. Wang, C. Gou, Y. Duan, Y. Lin, X. Zheng, and F. Y. Wang, “Generative adversarial networks: 1378
Introduction and outlook,” IEEE/CAA Journal of Automatica Sinica, vol. 4, no. 4, pp. 588–598, Oct. 1379
2017, doi: 10.1109/JAS.2017.7510583. 1380
[177] S. Nema, A. Dudhane, S. Murala, and S. Naidu, “RescueNet: An unpaired GAN for brain tumor 1381
segmentation,” Biomed Signal Process Control, vol. 55, p. 101641, Jan. 2020, doi: 1382
10.1016/J.BSPC.2019.101641. 1383
[178] G. Neelima, D. R. Chigurukota, B. Maram, and B. Girirajan, “Optimal DeepMRSeg based tumor 1384
segmentation with GAN for brain tumor classification,” Biomed Signal Process Control, vol. 74, p. 1385
103537, Apr. 2022, doi: 10.1016/J.BSPC.2022.103537. 1386
[179] M. Rezaei, H. Yang, and C. Meinel, “Voxel-GAN: Adversarial framework for learning imbalanced 1387
brain tumor segmentation,” Lecture Notes in Computer Science (including subseries Lecture Notes 1388
in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11384 LNCS, pp. 321–333, 1389
2019, doi: 10.1007/978-3-030-11726-9_29/COVER/. 1390
[180] Y. Hua, J. Guo, and H. Zhao, “Deep Belief Networks and deep learning,” Proceedings of 2015 1391
International Conference on Intelligent Computing and Internet of Things, ICIT 2015, pp. 1–4, May 1392
2015, doi: 10.1109/ICAIOT.2015.7111524. 1393
[181] A. Kharrat and M. Néji, “Classification of brain tumors using personalized deep belief networks on 1394
MRImages: PDBN-MRI,” https://doi.org/10.1117/12.2522848, vol. 11041, pp. 713–721, Mar. 1395
2019, doi: 10.1117/12.2522848. 1396
[182] A. Ratna Raju, S. Pabboju, and R. Rajeswara Rao, “Hybrid active contour model and deep belief 1397
network based approach for brain tumor segmentation and classification,” Sensor Review, vol. 39, 1398
no. 4, pp. 473–487, Jul. 2019, doi: 10.1108/SR-01-2018-0008/FULL/XML. 1399
[183] M. Tschannen, O. Bachem, and M. Lucic, “Recent Advances in Autoencoder-Based Representation 1400
Learning,” Dec. 2018, doi: 10.48550/arxiv.1812.05069. 1401
[184] M. Tschannen, O. Bachem, and M. Lucic, “Recent Advances in Autoencoder-Based Representation 1402
Learning,” Dec. 2018, doi: 10.48550/arxiv.1812.05069. 1403
[185] J. Amin et al., “Brain Tumor Detection by Using Stacked Autoencoders in Deep Learning,” Journal 1404
of Medical Systems 2019 44:2, vol. 44, no. 2, pp. 1–12, Dec. 2019, doi: 10.1007/S10916-019-1483-1405
2. 1406
[186] M. M. Badža and M. Barjaktarović, “Segmentation of Brain Tumors from MRI Images Using 1407
Convolutional Autoencoder,” Applied Sciences 2021, Vol. 11, Page 4317, vol. 11, no. 9, p. 4317, 1408
May 2021, doi: 10.3390/APP11094317. 1409
[187] J. N. Stember and H. Shalu, “Deep Reinforcement Learning with Automated Label Extraction from 1410
Clinical Reports Accurately Classifies 3D MRI Brain Volumes,” Journal of Digital Imaging 2022, 1411
pp. 1–10, May 2022, doi: 10.1007/S10278-022-00644-5. 1412
[188] R. Dey and F. M. Salemt, “Gate-variants of Gated Recurrent Unit (GRU) neural networks,” Midwest 1413
Symposium on Circuits and Systems, vol. 2017-Augus, pp. 1597–1600, Sep. 2017, doi: 1414
10.1109/MWSCAS.2017.8053243. 1415
[189] A. M. G. Allah, A. M. Sarhan, and N. M. Elshennawy, “Classification of Brain MRI Tumor Images 1416
Based on Deep Learning PGGAN Augmentation,” Diagnostics 2021, Vol. 11, Page 2343, vol. 11, 1417
no. 12, p. 2343, Dec. 2021, doi: 10.3390/DIAGNOSTICS11122343. 1418
[190] J. Chang et al., “A mix-pooling CNN architecture with FCRF for brain tumor segmentation,” J Vis 1419
Commun Image Represent, vol. 58, pp. 316–322, Jan. 2019, doi: 10.1016/J.JVCIR.2018.11.047. 1420
[191] H. N. T. K. Kaldera, S. R. Gunasekara, and M. B. DIssanayake, “Brain tumor Classification and 1421
Segmentation using Faster R-CNN,” 2019 Advances in Science and Engineering Technology 1422
International Conferences, ASET 2019, May 2019, doi: 10.1109/ICASET.2019.8714263. 1423
[192] M. Sajjad, S. Khan, K. Muhammad, W. Wu, A. Ullah, and S. W. Baik, “Multi-grade brain tumor 1424
classification using deep CNN with extensive data augmentation,” J Comput Sci, vol. 30, pp. 174–1425
182, Jan. 2019, doi: 10.1016/J.JOCS.2018.12.003. 1426
[193] M. Y. B. Murthy, A. Koteswararao, and M. S. Babu, “Adaptive fuzzy deformable fusion and 1427
optimized CNN with ensemble classification for automated brain tumor diagnosis,” Biomedical 1428
Engineering Letters 2021 12:1, vol. 12, no. 1, pp. 37–58, Nov. 2021, doi: 10.1007/S13534-021-1429
00209-5. 1430
[194] A. Rehman, M. A. Khan, T. Saba, Z. Mehmood, U. Tariq, and N. Ayesha, “Microscopic brain tumor 1431
detection and classification using 3D CNN and feature selection architecture,” Microsc Res Tech, 1432
vol. 84, no. 1, pp. 133–149, Jan. 2021, doi: 10.1002/JEMT.23597. 1433
[195] J. Amin, M. Sharif, M. Raza, T. Saba, R. Sial, and S. A. Shad, “Brain tumor detection: a long short-1434
term memory (LSTM)-based learning model,” Neural Computing and Applications 2019 32:20, vol. 1435
32, no. 20, pp. 15965–15973, Dec. 2019, doi: 10.1007/S00521-019-04650-7. 1436
[196] G. Liu, X. Li, and Y. Cai, “Segmentation for Multimodal Brain Tumor Images Using Dual-Tree 1437
Complex Wavelet Transform and Deep Reinforcement Learning,” Comput Intell Neurosci, vol. 1438
2022, pp. 1–11, May 2022, doi: 10.1155/2022/5369516. 1439
[197] T. Sathies Kumar, C. Arun, and P. Ezhumalai, “An approach for brain tumor detection using optimal 1440
feature selection and optimized deep belief network,” Biomed Signal Process Control, vol. 73, p. 1441
103440, Mar. 2022, doi: 10.1016/J.BSPC.2021.103440. 1442
[198] P. Harish and S. Baskar, “MRI based detection and classification of brain tumor using enhanced 1443
faster R-CNN and Alex Net model,” Mater Today Proc, Dec. 2020, doi: 1444
10.1016/J.MATPR.2020.11.495. 1445
[199] B. Ahmad, J. Sun, Q. You, V. Palade, and Z. Mao, “Brain Tumor Classification Using a 1446
Combination of Variational Autoencoders and Generative Adversarial Networks,” Biomedicines 1447
2022, Vol. 10, Page 223, vol. 10, no. 2, p. 223, Jan. 2022, doi: 10.3390/BIOMEDICINES10020223. 1448
[200] D. Mukherkjee, P. Saha, D. Kaplun, A. Sinitca, and R. Sarkar, “Brain tumor image generation using 1449
an aggregation of GAN models with style transfer,” Scientific Reports 2022 12:1, vol. 12, no. 1, pp. 1450
1–16, Jun. 2022, doi: 10.1038/s41598-022-12646-y. 1451
[201] W. Takrouni and A. Douik, “Improving geometric P-norm-based glioma segmentation through deep 1452
convolutional autoencoder encapsulation,” Biomed Signal Process Control, vol. 71, p. 103232, Jan. 1453
2022, doi: 10.1016/J.BSPC.2021.103232. 1454
[202] A. Chattopadhyay and M. Maitra, “MRI-based brain tumour image detection using CNN based deep 1455
learning method,” Neuroscience Informatics, vol. 2, no. 4, p. 100060, Dec. 2022, doi: 1456
10.1016/J.NEURI.2022.100060. 1457
[203] N. Kesav and M. G. Jibukumar, “Efficient and low complex architecture for detection and 1458
classification of Brain Tumor using RCNN with Two Channel CNN,” Journal of King Saud 1459
University - Computer and Information Sciences, May 2021, doi: 10.1016/J.JKSUCI.2021.05.008. 1460
[204] R. Vankdothu, M. A. Hameed, and H. Fatima, “A Brain Tumor Identification and Classification 1461
Using Deep Learning based on CNN-LSTM Method,” Computers and Electrical Engineering, vol. 1462
101, p. 107960, Jul. 2022, doi: 10.1016/J.COMPELECENG.2022.107960. 1463
[205] L. Fidon, S. Ourselin, and T. Vercauteren, “Generalized Wasserstein Dice Score, Distributionally 1464
Robust Deep Learning, and Ranger for Brain Tumor Segmentation: BraTS 2020 Challenge,” 1465
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and 1466
Lecture Notes in Bioinformatics), vol. 12659 LNCS, pp. 200–214, 2021, doi: 10.1007/978-3-030-1467
72087-2_18/COVER/. 1468
[206] F. Isensee, P. F. Jäger, P. M. Full, P. Vollmuth, and K. H. Maier-Hein, “nnU-Net for Brain Tumor 1469
Segmentation,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial 1470
Intelligence and Lecture Notes in Bioinformatics), vol. 12659 LNCS, pp. 118–132, 2021, doi: 1471
10.1007/978-3-030-72087-2_11/COVER/. 1472
[207] T. Henry et al., “Brain Tumor Segmentation with Self-ensembled, Deeply-Supervised 3D U-Net 1473
Neural Networks: A BraTS 2020 Challenge Solution,” Lecture Notes in Computer Science 1474
(including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 1475
vol. 12658 LNCS, pp. 327–339, 2021, doi: 10.1007/978-3-030-72084-1_30/COVER/. 1476
[208] Y. Yuan, “Automatic Brain Tumor Segmentation with Scale Attention Network,” Lecture Notes in 1477
Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in 1478
Bioinformatics), vol. 12658 LNCS, pp. 285–294, 2021, doi: 10.1007/978-3-030-72084-1479
1_26/COVER/. 1480
[209] J. Tang, T. Li, H. Shu, and H. Zhu, “Variational-Autoencoder Regularized 3D MultiResUNet for 1481
the BraTS 2020 Brain Tumor Segmentation,” Lecture Notes in Computer Science (including 1482
subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 12659 1483
LNCS, pp. 431–440, 2021, doi: 10.1007/978-3-030-72087-2_38/COVER/. 1484
[210] C. A. Silva, A. Pinto, S. Pereira, and A. Lopes, “Multi-stage Deep Layer Aggregation for Brain 1485
Tumor Segmentation,” Lecture Notes in Computer Science (including subseries Lecture Notes in 1486
Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 12659 LNCS, pp. 179–188, 2021, 1487
doi: 10.1007/978-3-030-72087-2_16/COVER/. 1488
[211] T. Tarasiewicz, M. Kawulok, and J. Nalepa, “Lightweight U-Nets for Brain Tumor Segmentation,” 1489
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and 1490
Lecture Notes in Bioinformatics), vol. 12659 LNCS, pp. 3–14, 2021, doi: 10.1007/978-3-030-1491
72087-2_1/COVER/. 1492
[212] S. Qamar, P. Ahmad, and L. Shen, “HI-Net: Hyperdense Inception 3D UNet for Brain Tumor 1493
Segmentation,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial 1494
Intelligence and Lecture Notes in Bioinformatics), vol. 12659 LNCS, pp. 50–57, 2021, doi: 1495
10.1007/978-3-030-72087-2_5/COVER/. 1496
[213] P. Ahmad, S. Qamar, L. Shen, and A. Saeed, “Context Aware 3D UNet for Brain Tumor 1497
Segmentation,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial 1498
Intelligence and Lecture Notes in Bioinformatics), vol. 12658 LNCS, pp. 207–218, 2021, doi: 1499
10.1007/978-3-030-72084-1_19/COVER/. 1500
[214] J. Colman, L. Zhang, W. Duan, and X. Ye, “DR-Unet104 for Multimodal MRI Brain Tumor 1501
Segmentation,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial 1502
Intelligence and Lecture Notes in Bioinformatics), vol. 12659 LNCS, pp. 410–419, 2021, doi: 1503
10.1007/978-3-030-72087-2_36/COVER/. 1504
[215] Z. Jiang, C. Ding, M. Liu, and D. Tao, “Two-stage cascaded u-net: 1st place solution to brats 1505
challenge 2019 segmentation task,” Lecture Notes in Computer Science (including subseries Lecture 1506
Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11992 LNCS, pp. 231–1507
241, 2020, doi: 10.1007/978-3-030-46640-4_22/COVER/. 1508
[216] A. Myronenko and A. Hatamizadeh, “Robust semantic segmentation of brain tumor regions from 1509
3D MRIs,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial 1510
Intelligence and Lecture Notes in Bioinformatics), vol. 11993 LNCS, pp. 82–89, 2020, doi: 1511
10.1007/978-3-030-46643-5_8/COVER/. 1512
[217] R. R. Agravat and M. S. Raval, “Brain tumor segmentation and survival prediction,” Lecture Notes 1513
in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes 1514
in Bioinformatics), vol. 11992 LNCS, pp. 338–348, 2020, doi: 10.1007/978-3-030-46640-1515
4_32/COVER/. 1516
[218] F. Wang, R. Jiang, L. Zheng, C. Meng, and B. Biswal, “3d u-net based brain tumor segmentation 1517
and survival days prediction,” Lecture Notes in Computer Science (including subseries Lecture 1518
Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11992 LNCS, pp. 131–1519
141, 2020, doi: 10.1007/978-3-030-46640-4_13/COVER/. 1520
[219] M. U. Saeed et al., “RMU-Net: A Novel Residual Mobile U-Net Model for Brain Tumor 1521
Segmentation from MR Images,” Electronics 2021, Vol. 10, Page 1962, vol. 10, no. 16, p. 1962, 1522
Aug. 2021, doi: 10.3390/ELECTRONICS10161962. 1523
[220] G. K. Murugesan et al., “Multidimensional and multiresolution ensemble networks for brain tumor 1524
segmentation,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial 1525
Intelligence and Lecture Notes in Bioinformatics), vol. 11993 LNCS, pp. 148–157, 2020,