Magnetic Resonance Imaging 61 (2019) 247–259

Contents lists available at ScienceDirect

Magnetic Resonance Imaging

journal homepage: www.elsevier.com/locate/mri

Review article

A review on brain tumor segmentation of MRI images T

a a,* b
Anjali Wadhwa , Anuj Bhardwaj , Vivek Singh Verma
a
Jaypee Institute of Information Technology, Noida, 201309, India
b
Ajay Kumar Garg Engineering College, Ghaziabad, 201009, India

ARTICLE INFO ABSTRACT

Keywords: The process of segmenting tumor from MRI image of a brain is one of the highly focused areas in the community
MRI of medical science as MRI is noninvasive imaging. This paper discusses a thorough literature review of recent
Segmentation methods of brain tumor segmentation from brain MRI images. It includes the performance and quantitative
Brain tumor analysis of state-of-the-art methods. Different methods of image segmentation are briefly explained with the
Classification
recent contribution of various researchers. Here, an effort is made to open new dimensions for readers to explore
Ensemble learning
the concerned area of research. Through the entire review process, it has been observed that the combination of
Conditional Random Field (CRF) with Fully Convolutional Neural Network (FCNN) and CRF with DeepMedic or
Ensemble are more effective for the segmentation of tumor from the brain MRI images.

1. Introduction modalities like MRI, CT, and PET [6]. The advancement of medical
imaging devices especially MRI provides comprehensive analysis and
Image segmentation plays a crucial role in image understanding, extract pertinent information from images of most human body struc-
feature extraction, analysis and interpretation for many applications. It tures. A significant development in computational techniques helps the
has widespread applications in medical science, for example, tissue specialists to investigate the tremendous measure of information con-
classification, localization of tumors, tumor volume estimation, deli- tained in medical images [2]. Imaging techniques like MRI are used to
neation of blood cells, surgical planning, atlas matching, and image find the position and degree of extent of brain tumors. MRI provides
registration [1]. Mathematical algorithms of feature extraction, mod- noteworthy contrast for different brain tissues in comparison to other
eling and measurement can be exploited in the images to detect pa- imaging modalities like CT [7, 8]. Moreover, MR images typify sig-
thology, an evolution of the disease, or to compare a normal subject to nificant information regarding various tissue parameters (PD, spinlat-
an abnormal one [2]. The accurate and reproducible quantification and tice (T1) and spinspin (T2) relaxation times, flow velocity and chemical
morphology of tumors are of crucial importance for diagnosis, treat- shift), which prompt more precise brain tissue portrayal. T2 weighted
ment planning as well as monitoring of response to oncologic therapy (T2) images are generally utilized for giving an underlying evaluation,
for brain tumors [3]. Brain tumor segmentation consists of separating recognizing tumor types, and segregating tumors from non-tumor tis-
the different tumor tissues (active tumor, edema and necrosis) from sues. The contrast material in T1 weighted images (T1c) helps to en-
normal brain tissues: GM, WM, and CSF [4]. hance tumor boundaries from the neighbouring normal tissues [9]. The
A brain tumor is a cancerous or non-cancerous growth of abnormal T2 weighted scan in axial viewing with FLAIR is used to show non-
cells in the brain. A brain tumor can be benign or malignant. The benign enhanced tumors [10]. In the light of these special features, MRI gives
brain tumor has a uniformity in structure and does not contain active an edge for decision in brain tumor studies [11].
(cancer) cells, whereas malignant brain tumors have a non-uniformity Brain MRI segmentation is an essential basic step that has many
in structure and contain active cells. The gliomas and meningiomas are applications in neurology such as quantitative analysis, operational
the examples of low-grade tumors, classified as benign tumors. They planning, and functional imaging [1]. Although MRI can describe the
look like normal brain cells and grow slowly. Glioblastoma and astro- structures of the brain accurately, however medical image segmenta-
cytomas are a class of high-grade tumors, classified as malignant tu- tion is a tough task because of poor spatial resolution, low contrast, ill-
mors [5]. They tend to grow rapidly. defined boundaries, inhomogeneity, partial volume effect, noise,
Clinicians plan the most responsive and viable treatment for pa- variability of object shapes and some other acquisition artifacts in the
tients by acquiring the data from various restorative diagnostic imaging retrieved data as well as the lack of models of the anatomy that entirely

*
Corresponding author.
E-mail address: [email protected] (A. Bhardwaj).

https://doi.org/10.1016/j.mri.2019.05.043
Received 13 January 2019; Received in revised form 30 May 2019; Accepted 30 May 2019
0730-725X/ © 2019 Elsevier Inc. All rights reserved.
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

capture the possible deformations in each structure [1, 12, 13]. These Table 1
limitations can be dealt with by implementing few simple approaches List of abbreviations.
as proposed in Refs. [13-16]. The heterogeneous appearance of tumors Description Abbreviation
incorporating the expansive changeability in location, size, shape and
recurrence makes it difficult to devise effective segmentation rules [12]. Active Contour Models ACM
AdaBoost Algorithm with Random Forests ADBRF
Notwithstanding these issues, a notable success is recorded in this field
Adaptive Fuzzy Inference System ANFIS
as a result of continuing advancement in instrumentation and computer Artificial Bee Colony ABC
technology. Artificial Bee Colony-Standard Particle Swarm Optimization ABC-SPSO
Many researchers have classified schemes of image segmentation Artificial Neural Network ANN
under different heads. One of such classifications as discussed in Refs. Berkeley Wavelet Transform BWT
Boundary Vector Flow BVF
[4, 6] are threshold-based, region-based, pixel classification and model-
Cerebrospinal Fluid CSF
based techniques. The other such classifications as presented in Refs. Computerized Tomography CT
[17-19] are threshold-based, boundary-based and region-based Conditional Random Field CRF
methods. Threshold-based methods are based on the postulate that the Conditional Random Fields - Recurrent Neural Network CRF-RNN
Content-Based Active Contour CBAC
pixels that lie within a certain range belong to one class [20, 21].
Convolutional Neural Network CNN
Boundary-based methods use the assumption that pixel properties Darwinian Particle Swarm Optimization DPSO
change abruptly at the boundary of two regions [2, 22-33]. Region- Deep Neural Network DNN
based methods assume that a region is composed of adjacent pixels with Dice Similarity Coefficient DSC
similar properties [17, 34-36]. In pixel classification techniques [37, Directional Spectral Distribution DSD
Diffusion Tensor Image Segmentation D-SEG
38], segmentation depends upon feature space using pixel attributes
Diffusion Tensor Imaging DTI
that may consist of gray level, local texture and color components for Discrete Wavelet Transform DWT
each pixel in the image. In model-based techniques as in Refs. [22, 39- Edge Stop Function ESF
41], a model is formed incorporating a-priori knowledge such as shape, Enhanced Fuzzy C-Means Algorithm EnFCM
size, texture, orientation about the object or specific anatomical struc- Fast Generalized Fuzzy C-Means FGFCM
Feed-forward Neural Network FNN
ture. There are some hybrid approaches that concatenate two or more Fluid-Attenuated Inversion Recovery FLAIR
of the aforementioned methods [42-45]. Fluid Vector Flow FVF
Depending upon the degree of human interaction required in seg- Fractional Anisotropy FA
mentation of an image, segmentation techniques are classified as Fractional Order Darwinian Particle Swarm Optimization FODPSO
Fully Convolutional Neural Network FCNN
manual, semi-automatic [34, 46-48] and fully automatic [1, 3, 9, 42,
Fuzzy C-Means FCM
49, 50]. In manual segmentation, the clinicians locate tumors on their Fuzzy local information C-Means FLICM
own using their experience and knowledge of anatomy, thus consuming Gaussian Mixture Models GMM
a lot of time. The results through manual segmentation vary from Genetic Algorithm GA
person to person as it relies on the knowledge of operator and also Glioblastoma Multiforme GBM
Gradient Vector Flow GVF
requires a lot of expertise. It is regarded as ground truth for semi-au- Gray Matter GM
tomatic and automatic methods. Semi-automatic methods combine Gray-Level Co-Occurrence Matrix GLCM
computers and human expertise whereas in fully automatic methods, Hierarchical Centroid Shape Descriptor HCSD
computers determine the segmentation without any human interven- High-Grade Gliomas HGG
Hybridization of PSO and ABC HPA
tion. The former requires human intervention in the form of in-
Integrated Algorithm Based on ABC and PSO IABAP
itialization or manually correcting or modifying the outcomes which Intensity Inhomogeneity IIH
results in efficient segmentation; however, user intervention may tend k-Nearest Neighbour k-NN
segmentation results to suffer variations from person to person and Leaky Rectified Linear Unit LReLU
even within the same user. The latter requires anatomical knowledge Least Squares Support Vector Machines LSVM
Level Set Function LSF
like the shape, size, appearance and location of tumor for construction Levenberg-Marquardt LM
of a model and performing the task. Moreover, it should be able to Low-Grade Gliomas LGG
adapt to variations in the characteristics which makes it challen- Magnetic Resonance Imaging MRI
ging [4]. Clinically, semi-automatic methods with least user interven- Modified Region Growing Method MRGM
Multi-Fractal Analysis MFA
tion are more acceptable.
Multi-modal Magnetic Resonance Imaging mMRI
Depending on the use of manually labelled training data, segmen- Multi-parametric Magnetic Resonance Imaging mpMRI
tation methods are classified as supervised [5, 51, 52] and unsupervised Multi-Population Cuckoo Search Strategy MCSS
methods [53-55]. Supervised methods utilize the labelled data in the Multi-Population Particle Swarm Optimization Strategy MPSOS
construction of a model that maps the extracted features to labels or Neighbourhood Intuitionistic FCM Clustering Algorithm with GA NIFCMGA
Particle Swarm Optimization PSO
classes in the training phase which are then used to determine the Pizza-Slice Shaped Search Windows PSSW
classes of the unlabelled data in the testing phase. The training phase Positron-Emission tomography PET
requires human intervention which ends up in variability of results. The Principal Component Analysis PCA
reference is made to understand the details of the supervised learning Probabilistic Neural Network PNN
Proton Density PD
based method proposed by Raschke et al. [56]. In unsupervised seg-
Radial Basis Function RBF
mentation, no training data is available to group pixels with homo- Random Forest RF
geneous attributes together. The number of classes are automatically Rectangular Shaped Search Windows RSSW
decided by an algorithm, making the use of features like intensity and Recurrent Neural Network RNN
texture. These methods are useful to handle more complicated cases. Region of Interest ROI
Selective Binary Gaussian Filtering Regularizing Level Set SBGFRLS
It is impossible to develop only one standard method that suits all Signal-to-Noise Ratio SNR
imaging applications. It is in this context that an overview of the Stationary Wavelet Transform SWT
methodologies of image segmentation is presented. This paper focuses Support Vector Machines SVM
on the brain tumor segmentation of data obtained from MRI using Volume of Interest VOI
White Matter WM
different techniques for segmentation. Reference is made to Table 1 for
abbreviations used throughout in this review work.

248
 A. Wadhwa, et al.

Table 2
Overview of recent methods on the basis of different features.
References MRI modalities Method Time (s) Dim Supervision Type of abnormality

[108] T1c,T2,FLAIR FCNN + CRF-RNN – 2D – HGG, LGG

[51] T1, T1c, T2, FLAIR Incremental CNN 21 (Approx.) 2D Automatic High and low grade glioblastomas
[52] T1c, T2, FLAIR, isotropic (p) and anisotropic (q) Supervoxel based random forests – 3D – All grade tumors
components derived from the diffusion tensor imaging
(DTI)
[12] T1, T1c, T2, FLAIR CNN + fully connected CRF (DeepMedic + CRF) < 180 3D Automatic Traumatic brain injuries, brain tumors (HGG, LGG) and
ischemic stroke
[105] T1, T1c, T2, FLAIR DNN 25 to 180 2D Automatic High and low grade glioblastomas
[83] T2 FODPSO-DSDMSA < 5 2D Automatic Gliomas
[5] T1, T2, FLAIR, PD BWT + SVM – 2D – Glial Tumors
[50] FLAIR Superpixel based extremely randomized trees – 2D Automatic HGG, LGG

249
[54] T1, T1c SBGFRLS – 2D Semi-automatic Homogeneously-enhanced, non-enhanced,
heterogeneously-enhanced and ring- enhanced tumor
[149] T1, T1c, T2, FLAIR LACM-BIC 448.57 s per patient and 2D Semi-automatic All grade tumors
15.82 s per image slice
[150] FLAIR, T1, T1c, T2 FCNN + Texton + RF – 2D Automatic HGG, LGG
[104] FLAIR, T1, T1c, T2 U-Net based DNN 2 to 3 2D Automatic HGG, LGG
[100] T1, T1c, T2, FLAIR CNN – 2D Automatic HGG, LGG
[90] T2 DWT + PPCA + ADBRF 0.0243 2D Automatic Gliomas
[9] T1, T1c, T2, FLAIR Three-dimensional active contour without edge 243 s per slice 3D Automatic various types
[85] T1, T1c, T2, FLAIR Symmetric Multimodal Templates + Concatenated – 2D Automatic HGG, LGG
Random Forests
[87] T1, T2, FLAIR GMM feature extraction + NB/SVM/PNN classifier – 2D Automatic Glioblastoma
[86] T1, T1c, T2, FLAIR Extremely Randomized Forest with High-Level – 2D Automatic Gliomas
Features
[99] T1, T1c, T2, FLAIR Deep CNN 480 2D Semi-automatic HGG, LGG
Magnetic Resonance Imaging 61 (2019) 247–259
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

Table 3
Comparative study of recent schemes on BRATS 2015 data sets.
Methods Dice Precision Sensitivity

Comp. Core Enh. Comp. Core Enh. Comp. Core Enh.

110 cases [108]

0.84 0.73 0.62 0.89 0.76 0.63 0.82 0.76 0.67
DeepMedic + CRF in Ref. [12] 0.85 0.67 0.63 0.85 0.85 0.63 0.88 0.61 0.66
Ensemble + CRF in Ref. [12] 0.85 0.67 0.63 0.85 0.86 0.63 0.88 0.60 0.67

274 cases
DeepMedic + CRF in Ref. [12] 0.90 0.75 0.72 0.93 0.84 0.76 0.90 0.72 0.73
Ensemble + CRF in Ref. [12] 0.90 0.75 0.73 0.92 0.86 0.76 0.89 0.72 0.74
[154] 0.88 0.77 0.68 0.90 0.84 0.68 0.89 0.76 0.75
[99] 0.87 0.73 0.68 0.89 0.74 0.72 0.86 0.77 0.70

*Comp. — Complete, Enh. — Enhancing, Bold faces for good performance.

Table 4
Comparative study of recent schemes on BRATS 2013.
Methods Dice Precision Sensitivity

Comp. Core Enh. Comp. Core Enh. Comp. Core Enh.

Challenge
Zhao et al. [108] 0.88 0.84 0.77 0.90 0.87 0.76 0.86 0.82 0.80
Pereira et al. [100] 0.88 0.83 0.77 0.88 0.87 0.74 0.89 0.83 0.81
Havaei et al. [105] 0.86 0.77 0.73 0.88 0.85 0.76 0.78 0.68 0.58
Tustison et al. [85] 0.87 0.78 0.74 0.85 0.74 0.69 0.89 0.88 0.83
Kwon et al. [155] 0.88 0.83 0.72 0.92 0.90 0.74 0.84 0.78 0.72
Davy et al. [156] 0.85 0.74 0.68 0.85 0.74 0.62 0.85 0.78 0.77

Leaderboard
Zhao et al. [108] 0.86 0.73 0.62 0.89 0.77 0.60 0.85 0.77 0.69
Pereira et al. [100] 0.84 0.72 0.62 0.85 0.82 0.60 0.86 0.76 0.68
Havaei et al. [105] 0.83 0.69 0.59 0.86 0.78 0.55 0.84 0.71 0.67
Tustison et al. [85] 0.79 0.65 0.53 0.83 0.7 0.51 0.81 0.73 0.66
Kwon et al. [155] 0.86 0.79 0.59 0.88 0.84 0.60 0.86 0.81 0.63
Davy et al. [156] 0.72 0.63 0.56 0.69 0.64 0.5 0.82 0.68 0.68

*Comp. — Complete, Enh. — Enhancing, Bold faces for good performance.

1.1. Key contributions 2. Methods of segmentation

The present work is devoted to providing an abstract idea, parti- 2.1. Conventional methods
cularly, the analysis of the existing methods that are used to segment
the tumor present in MRI images. A number of methods are available to 2.1.1. Thresholding
segment the tumor from brain MRI images. Although, the survey Thresholding is one of the simplest techniques of image segmenta-
paper [4] provides information on different methods of segmentation tion. It is fast, easy to implement and understand. It works upon the
with their advantages and disadvantages, there is still scope not only for idea of converting a scalar image into a binary image, in which a
comparative analysis but also for the contribution of recent methods. threshold value is decided based upon the intensity values of the image.
The highlights of this study are as follows: The intensity values of the pixels are compared with the threshold
value. The pixels with intensity value same-as or higher than threshold
1. It presents an exhaustive study of recent methods for brain tumor value are assigned the value 1, whereas the pixels with lower intensity
segmentation from brain MRI images. values are assigned 0, thus separating the foreground (white pixels) and
2. It helps the clinicians to take the decision of correct diagnosis and the background region (black pixels), respectively.
further treatment accordingly. Consider the original gray scale image, say f(i,j). Depending upon
3. Quantitative analysis through different metrics shows the effec- the intensity values, an initial threshold value T is chosen. Partition the
tiveness and appropriateness of recent schemes. image into two sets H1 and H2 where H1 and H2 contains the set of
4. It also gives readers new directions of research for brain tumor pixels that are brighter and darker than the threshold value respec-
segmentation. tively. Further, the mean intensities h1 and h2 of H1 and H2 respectively
are evaluated depending upon which a new threshold value is calcu-
The remaining paper is organized as: Section 2 presents various lated as T1 = 1 2 2 . If ∥T − T1∥≥ΔT, (a pre-decided parameter), then
h +h

segmentation methods used in literature, Section 3 focuses on com- the above procedure is repeated else the binary image h(i,j) is obtained
parison and discussions. Finally, Section 4 depicts the conclusions. as

1 if f (i, j) TF
h (i , j ) =
0 if f (i , j ) < TF (1)

where TF is the final threshold value. This is an iterative way of se-

lecting a threshold as used in Refs. [57, 58].
When a single threshold value is chosen for the entire image, it is

250
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

called a global threshold. This is the most intuitive approach for Fabijanska [65] in seeded region growing method for tumor segmen-
thresholding of an image. It is computationally easy and fast. It does not tation along with some pre-processing, yielding satisfactory results.
incorporate any local relationship between pixels. It does not require However, the adaptive threshold may prove to be more reliable. In spite
any prior shape knowledge so it is appropriate for segmentation of of considering only intensity constraint, inclusion of orientation con-
images that do not have any fixed shape. Otsu's method [20] quests for straint in Ref. [66] has lead to improved results. An enhanced version of
finding the optimum value of the global threshold to segregate object traditional region growing has been developed in Ref. [36]. A brain
from the background in the image. The histogram is assumed to be MRI is processed from generation of threshold T2 and PD image of a
bimodal in Otsu's method. This method fails if two classes are of dif- brain MRI using seeded region growing algorithm, which is further
ferent sizes or with variable illumination across the image [59]. Sujan processed by Markov Logic Algorithm leading to classification based on
et al. [21] have used Otsu's thresholding along with morphological tumor's existence.
operations like dilation and erosion to detect brain tumor from MRI
image. A new threshold method is presented in Ref. [60] to produce 2.2. Supervised methods
more accurate results. The authors in Ref. [61] have attempted to find a
global threshold using level sets for segmentation of tumor and non- Supervised methods utilize the accurately labelled data obtained in
tumor regions. It requires only zero level set to be fed into it to perform the training phase to decide classes for unlabelled data of testing
the task. But, its efficiency becomes questionable if the difference of the phase [4]. It involves two phases: training and testing. In the training
intensity levels of tumorous and non-tumorous regions is lower. The phase, a model is constructed that maps the extracted features of data
performance of global thresholding approach deteriorates if the in- points to labels or classes. The model is then used to determine the
tensity of image pixels is of low contrast and not homogenous or in high classes of the unlabelled data in the testing phase. The training phase
noise levels. demands human intervention which ends up demonstrating the varia-
It is not always possible to divide an image into two regions by using bility of results. Supervised classification performs better than un-
a single threshold value. An image may have more than two types of supervised classification in terms of classification accuracy. Few clas-
regions, wherein the objects in the image do not share same intensity sifiers are discussed as:
values. In such a scenario, multiple threshold values are used to seg-
ment an image into various regions of interest. To meet the requirement 2.2.1. k-Nearest Neighbour
in these situations, several thresholding methods are proposed in lit- k-NN algorithm is a memory-based supervised learning algorithm
erature as local and dynamic threshold [62]. These methods are suited that directly compares the new unlabelled problem instances with a
for segmentation when it is not possible to predict a single threshold collection of labelled samples in the training set. k-NN classification is
value from the histogram of the image. performed in two stages. In the first stage, nearest neighbours for an
Thresholding is mostly used as a pre-processing step to segmenta- unlabelled instance are identified and then in the second stage, the class
tion of complex images like those of brain MRI because of their in- or label of the instance is determined using those neighbours. The
capability to exploit all the relevant information from the image. training phase of the algorithm consists of storing the feature vectors
and class labels of the training samples. The training data is being
2.1.2. Region growing stored and then the similarity or distance between the new unlabelled
Region growing is a fundamental and a popular method of image test point and the training instances is computed to predict the label of
segmentation used to segment homogeneous regions that have the same the test point. k Nearest neighbours to the unlabelled data from the
intensity values. It does not require any prior shape knowledge so it can training set are identified and their labels are used to assign the label or
be implemented on any object that varies in shape. The main principles class to the unknown record by majority voting. The parameter k de-
of working include that every pixel must be in some region, pixels in a pends upon the number of extracted features and the number of
region must be connected and should satisfy some specific similarity cases [67]. The distance metric to be used depends upon the problem
conditions, regions should be disjoint, and two different regions must domain [68]. k-NN is used in Ref. [67] to segment light and dark ab-
not have the same properties [17]. It initializes with a seed point in normalities within both medium and low background gray level values
each region of interest that can be selected manually or automatically. in the FLAIR-MRI brain images. Significant improvement in the seg-
Then the neighbouring pixels or regions are connected to the seed point mentation of white matter lesions has been noticed in Ref. [69] due to
according to some predefined similarity criteria. The method is kept inclusion of tissue type priors in feature set and use of variance scaling
progressing till all the pixels are classified in one of the regions main- for intensity normalization.
taining the connectivity of all the pixels grown from the seed point. It
incorporates local relationships between pixels. Its computational
2.2.2. Support vector machines
simplicity enhances the use of this method; however, it is sensitive to
SVM splits the image into two classes by finding a hyperplane that
seed initialization and noise. It does not perform well for images with
best classifies the data [70]. For instance, the hyperplane which has the
neighbouring regions sharing similar intensities and non-smoothly
maximum distance from the nearest data point on each side is
varying regions like textured images. It can be applied to the images
chosen [71]. SVM requires solving the following optimization problem:
affected by lighting variation but only after suitable pre-processing.
Partial volume effect is also an important factor that limits the accuracy 1 T
l

of segmentation as it blurs the intensity distinction at the border of the min, w w+C i,
w, b, 2 i=1 (2)
two tissue types, thereby making a voxel to represent more than one
tissue type [35]. This is overcome by the introduction of MRGM [63] by subject to yi (wT (x i ) + b) 1 i , with ξi ≥ 0, where (xi,yi) are la-
using gradient information for identification of boundaries. A com- belled pairs in data set, w is the normal vector of separating hyperplane,
parative analysis of traditional region growing and MRGM is presented ξ is error term, C is penalty parameter for error term and b is the offset
in Ref. [34] which proves MRGM to give better results of tumor seg- of the hyperplane [72]. SVM uses a kernel function K(xi,xj) = ϕ(xi)Tϕ
mentation in 3D T1 MRI images. (xj) to transform the data into a higher dimensional feature space so
An effective choice of threshold for selecting similarity condition that the data can be linearly separated with a maximum margin [73].
between two regions plays an important role in efficient segmentation. Kharrat et al. [74] used spatial gray level dependence method to
This gap is filled by Deng et al. [64] by minimizing the variance be- extract the texture features which are then chosen by GA to be fed into
tween homogeneous regions and ximising the gradient along the SVM for classification of brain tissues into a normal, benign or malig-
boundary. A fixed threshold value is used by Weglinski and nant tumor. SVM is combined with hierarchical regularization using

251
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

CRFs in Ref. [73] to give much promising results in separating healthy vectors and each tree casts a unit vote for the most popular class at
and tumor tissue and then sub-classifying both regions into the CSF, input x”. In literature, several techniques (such as Refs. [50, 85-89]) are
WM, GM and necrotic, active, edema region respectively in a novel developed using the application of RF based classification. Tustison
hierarchical way. Zhang et al. [75] contributed in the literature by in- et al. [85] applied several features including intensity, geometry, and
troducing a novel feature selection algorithm fusing multi-input data. asymmetry from multiple modalities to a random forests classifier. The
However, this system is modeled based on 2D information to segment use of symmetric multivariate templates has enhanced its performance.
each pixel in 3D MRI volumes which limits its validation on large data The work presented by Chaddad [87] uses GMMs to identify GBM
sets. features using different individual modalities (T1 and T2 weighted
In SVM, points are classified based upon the proximity with separ- images, FLAIR MR images) separately. Multithresholding segmentation
ating hyperplane which takes longer computational time to solve linear with morphological operations of MR images has been used to detect
or quadratic problems so formed. In lieu of reducing the number of infected regions. GMM features demonstrated better performance re-
calculations, Wu et al. [76] emphasized on segmenting the image into sulting in early treatment of GBM. Koley et al. [89] implemented RF for
superpixels rather than processing on voxels. SVM is trained using pattern classification after the identification and extraction of tumor
features extracted by multi-level Gabor wavelet filter and the re- area. The tumor region is quantified with 86 features to develop the
lationship between the neighbouring superpixels is learned. The output training data set, which are then given as input to the classifier. The
from SVM and affinity model is combined with CRFs and structural ADBRF has been used in Ref. [90] which promises to give quite accu-
noise is removed, resulting in final segmentation of GBM. Recently, an rate results with only 13 discriminant features. Goetz et al. [88] pro-
effort reducing the classification error using SVM is presented by Sol- posed a new random forest-based method which uses domain adapta-
taninejad et al. [77]. Classification of tumor grades(such as II, III and tion to reduce sample selection errors introduced by the sparse
IV) is performed using the first and second order statistical features sampling. Extremely randomized trees [91] (extra trees classifier) are
extraction. Whereas, the ROI segmentation is incorporated using either used in Refs. [50, 86]. They are computationally efficient and reduce
manually or through a superpixel based approach. Enhanced version of the variance due to more randomization. Pinto et al. [86] includes high
MRI brain tumor grading is presented in Ref. [78]. level local features including intensities in all sequences and context-
An efficient method for brain tumor classification based on D-SEG is based features calculated over non-linear transformation of the images
proposed by Jones et al. [79]. This technique uses a combined appli- while Soltaninejad et al. [50] considered different types of features like
cation of k-means clustering and SVM. To delineate VOI of abnormal intensity statistics, textons, and curvature features. Instead of using
tissues, the k-means clustering approach with similar isotropic and voxels, supervoxels are formed in Ref. [52] using multimodal MRI to
anisotropic diffusion properties is applied. Whereas, SVM is used for reduce computations. Information from multimodal images is combined
classifying various types of tumors such as LGG, glioblastomas, me- to produce supervoxel boundaries across multiple image protocols.
tastases, and meningiomas. D-SEG is also used in Ref. [80] to identify Statistical and texton features provide improved performance for clas-
tumor core ROIs to further discriminate GBMs from solitary metastasis sification of brain tumor supervoxels using RFs.
using the selected 2-D shape features extracted from DTI data assuming
that the 2-D shape features are normally distributed. However, this is 2.2.4. Artificial neural networks
the limitation of presented work which is later suppressed by Yang Animals respond to the environment because of very small con-
et al. [72] with the use of 3D shape descriptors to improve the accuracy nected units called neurons in the nervous system. Like animals, arti-
in the discrimination between the two. ficial neural network imitates the behaviour of neurons and respond to
Like the standard SVM where data points are classified by finding a the surroundings to achieve an optimal state. The artificial neurons, like
single hyperplane, points can also be classified by assigning them to the the originals, process the input signals and then transmit them from one
closest of the two parallel lines which are pushed apart as far as pos- to another, thus traversing multiple intermediate hidden layers of
sible [81]. Vaishnavee and Amshakala [13] have used this proximal connections to get the relevant output at the final layer. The known
SVM classifier to solve a linear system of equations for generating a characteristics are fed into input nodes and mathematical operations
linear or non-linear classification. The classifier performs its task with are performed on them to get a final classification. The output of each
much more accuracy and in lesser time. Bahadure et al. [5] also used artificial neuron is a non-linear function of the sum of its inputs. Similar
SVM for tumor detection. The quality of the image is firstly enhanced to a person who learns from his experiences, the technique is trained to
using the modified sigmoid function. Then BWT is applied for effective evaluate the parameters involved in mathematical operations so that
segmentation of brain MR image after using suitable threshold opera- the errors at the final stage can be minimized. As the learning proceeds,
tion for skull stripping. Textural features are extracted using GLCM and the weights with which these artificial neurons are connected to each
fed into SVM that gave an accuracy of 96.51% in contrast with an ac- other in different layers also vary. These weights are found out by using
curacy of 90.54% without feature extraction. The results are compared optimization methods. As the size of the network becomes very large,
with the ANFIS [82], Back Propagation, and k-NN which gave an ac- the complexity of the technique increases and is time-consuming.
curacy of 90.04%, 85.57% and 87.06% respectively. Moreover, efficient training of the network requires a large number of
Lahmiri [83] has used PSO, DPSO and FODPSO to segment the images. However, its ability to model any arbitrary, non-trivial dis-
glioma affected area from the normal region in the image. DSD sig- tribution is of practical advantage. Different types of neural networks
natures from the segmented image are computed and its features are are being used currently [92, 93].
extracted by means of MFA and further, the classification is performed FNN classifier has been used in Refs. [66, 94-96]. The FNN is one of
by SVM. The results show the achievement of highest accuracy in the the simplest ANNs. In this network, the flow of information is uni-
case of FODPSO for segmentation. directional, from the input nodes, through the hidden nodes, and to the
output nodes. It can classify separable non-linear patterns and can ap-
2.2.3. Random forests proximate any arbitrary continuous function. Various algorithms like
Random forest, an ensemble learning scheme is a simple, supervised Back Propagation algorithm, GA, PSO and ABC are used to train FNN
classification technique which runs efficiently on large data sets, han- but they easily end up in giving local optimum values of the parameters.
dles thousands of input variables without variable deletion and esti- Moreover, they suffer a drawback of high computational costs. FNN and
mates important features for classification. It is relatively robust to RBF neural network have been used with the modified region growing
outliers and noise. Breiman [84] defined a random forest as “a classifier involving both intensity and orientation constraint in Ref. [66] and
consisting of a collection of tree structured classifiers {h(x,Θk),k = 1, gave satisfactory results.
…} where the Θk are independent identically distributed random El-Dahshan et al. proposed a fast and robust hybrid intelligent

252
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

machine learning technique in Ref. [42] for automatic detection of computational cost is its major disadvantage. Therefore, 2D-CNNs have
brain tumor. It is based on feedback pulse-coupled neural network for been widely adopted in the brain tumor segmentation methods. In
segmentation, DWT for feature extraction, the PCA for reducing the continuation of these methods, Dong at al. [104] contributed by pro-
dimensionality of the wavelet coefficients and the feed forward back- posing a fully automatic brain tumor segmentation method using U-net
propagation neural network to classify inputs into healthy or patholo- based deep convolution network. Here, it is shown that their results are
gically cases. The experiments showed the classification accuracy to be good for core region and comparable for the complete region of the
99%. Damodharan and Raghavan claim in Ref. [94] that FNN performs tumor.
better than k-NN classifier and Bayesian classifier, measured on the A new DNN is presented in Ref. [105]. Since, the distribution of
basis of sensitivity, specificity and accuracy graphs. Skull stripping and tumor labels is unbalanced, a two-phase training procedure has been
morphological operations are applied to remove the noise from the used to train CNNs efficiently. The authors have explored two types of
input image after which thresholding is used to segment the con- architecture: two-pathway and cascaded architecture. Two-pathway
stituents of the brain; CSF, WM, GM. Statistic features like mean, var- architecture models both local features and global contextual features
iance, entropy and energy functions are calculated and fed into the simultaneously. Cascaded architecture took into consideration the de-
classifier to get the segmentation of the abnormal part from the normal pendence of labels of pixels. Two CNNs are stacked wherein the seg-
one. mentation outputs from one CNN are taken as input to the following
Wang et al. have used SWT to extract features from brain MRI CNN which has helped to achieve higher performance.
images in Ref. [95]. SWT helps in better extraction of features resulting CNNs are powerful, features are not required to be fed into the
in improved classification [97]. The authors have then used PCA for system, rather it extracts the required information for classification
feature reduction to reduce the dimensionality and hence, the compu- from the provided data [106]. However, few loopholes, including the
tational cost. FNN is trained by three different algorithms: an IABAP, amount of hyper-parameters, requirement of a large amount of training
ABC-SPSO and HPA. The classification results are shown to be excellent data that demands memory and a lot of computations that too increase
in terms of sensitivity, specificity, precision, and accuracy when FNN is further with deeper networks, discourage the employment of CNNs on
trained with HPA. medical imaging data [107].
Shree and Kumar [98] have focused on removing noise and Zhao et al. [108] have used image patches and image slices to train
smoothening the image that has improved SNR. In this scheme, the FCNNs and CRF-RNN, respectively. The three segmentation models are
DWT for extracting wavelet coefficients and GLCM for statistical feature trained using patches and slices in axial, coronal and sagittal views.
extraction are used after region growing segmentation to improve the Image patches have been used in training to handle millions or billions
performance and to reduce the complexity. This was followed by of parameters involved in deep models. Three results are obtained
morphological operations which remove the noise that can be formed corresponding to each view which are then combined using voting
after segmentation. The PNN classifier also implemented in Ref. [7] was based fusion strategy. Fusion results have shown improvement. An
then used to train and test the performance accuracy in the detection of improved variant of nonparametric nonuniform intensity normalization
tumor location in brain MRI images. The proposed technique is proved (N3) algorithm, denominated N4ITK, for bias field correction and in-
to be effective since the experimental results recorded nearly 100% and tensity normalization method as proposed in Ref. [91] has been used as
95% accuracy in training and testing data sets respectively while a pre-processing step. Thresholding is used as a post-processing step to
identifying normal and abnormal tissues from brain MR images. correct the false labels. The authors have evaluated the positive effects
S. Pereira et al. [99, 100] introduced an automatic method for of the pre and post processing steps in the process, but at the same time,
segmentation into five classes: necrosis, edema, non-enhancing, en- they add to the implementation time of segmentation. The authors also
hancing tumor and normal tissue. For the evaluation, three tumor re- witnessed that the results obtained by using 3 modalities T1c, T2,
gions; Enhancing tumor, Core (necrosis + non-enhancing tumor + en- FLAIR are highly competent with those obtained by using all the 4
hancing tumor) and the Complete tumor (all tumor classes) are modalities. So, T1 is proved to be almost redundant which can result in
considered. The process is automatic except for the initial stage in reduction of storage and data acquisition cost.
which the user needs to identify the grade of glioma manually. The In order to reduce the inference time, a sequential CNN architecture
authors have used the standardizing method proposed by Nyul has been implemented in spite of parallel CNN architecture in Ref. [51].
et al. [101] as a pre-processing step so that similar intensities depict the The proposed Incremental XCNet develops and trains CNN models si-
same tissue type in an image. Two CNN architectures, one for LGG and multaneously rather than working on conventional trial and error
another (deeper than LGG) for HGG are trained wherein LReLU acti- technique. It incorporates ensemble learning. It incorporates no post-
vation function is applied to all the layers except the last where softmax processing and least pre-processing steps.
is used. The use of small kernels along with more convolutional layers
has been proposed which requires a smaller number of weights to be 2.3. Unsupervised methods
trained. Moreover, regularization and data augmentation are used to
reduce overfitting. DSC metric for two data sets BRATS 2013 and Unsupervised segmentation requires no training data to group pixels
BRATS 2015 for the complete, core, enhancing regions are 0.88, 0.83, with homogeneous attributes together. The number of classes is auto-
0.77 and 0.78, 0.65, 0.75, respectively. matically decided by an algorithm, making the use of image-based
Intensity distribution in an image for various tissue types does not features like intensity, gradient or texture. These methods can tackle
allow adaptation to variations in scanners or individuals or time. To more complex problems. The unsupervised methods of segmentation
overcome this problem, Zhang et al. [102] have used multi-modality are discussed as:
information from T1, T2, and FA images as input to deep CNN for
segmenting isointense stage brain tissues. A 3D deep CNN has been used 2.3.1. Clustering
in Ref. [103] that can handle an arbitrary number of modalities. A 3D Clustering is an unsupervised pixel-based method for image seg-
CNN with 11 layers and two parallel pathways has been presented in mentation that partitions an unlabelled image data into clusters in
Ref. [12] which incorporates local and contextual information thus, which the pixels share some common characteristics. It is of two types:
resulting in improved segmentation results. 3D fully connected CRFs hard clustering and soft clustering. In hard clustering, a data point ei-
are used as a post-processing step that helps to get better prediction ther belongs to a cluster or not. Clusters do not overlap as in k-means
results. It helps to overcome computational burden while processing 3D clustering whereas in soft clustering, in spite of putting a data point into
medical scans. Although, a 3D information of the MRI data is an ad- a cluster completely, the probability of belonging to a cluster is as-
vantage to 3D-CNNs but an increase in the network size and signed to it. Clusters can overlap. A single data can belong to two or

253
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

more clusters at the same time as in Fuzzy C-Means clustering. 1. 0 ≤ uij ≤ 1,forall i, j,
k
2. j = 1 uij = 1, for all i ,
2.3.1.1. k-Means clustering. 3.
n
u < n, for all j.
i = 1 ij
k-Means clustering is a key technique in pixel-based methods [38].
Segmentation is done using k-means algorithm in following few steps. The larger membership value signifies higher propinquity to belong
Initially, k means m1(1),m2(1),…,mk(1) are considered corresponding to to the particular cluster. One piece of data may belong to two or more
the k clusters of the image. Then, the distance between each clusters. The initial membership values are assigned to each data point
observation xi, i = 1,2,…,n where each xi is a d-dimensional real xi to belong to a cluster j. Then, the centres of the clusters and the
feature vector and the means mj(t), j = 1,2,…,k for each cluster at tth membership values are updated. Now, the fuzzy partitioning is done
iteration, is calculated. An observation says x1 belongs to that cluster through an iterative minimization of the objective function. The process
whose mean has the least distance with x1. So, a cluster of observations terminates when the membership values in consecutive iterations are
is formed at each iteration and new means are calculated. The above close enough.
steps are repeated until the new means do not change further. Hence, The major disadvantages of the standard FCM algorithm are that it
the image is finally partitioned into k clusters. uses the non-robust Euclidean distance and it is sensitive to noise. The
k-Means algorithm, though simple and computationally efficient, former problem is dealt in Refs. [119, 120] by adopting some robust
may not give an optimum value even after a large number of iterations. distance measures such as Lp norms (0 < p ≤ 1) that helped to reduce
The use of a different distance function other than Euclidean distance the effect of outliers on results. In addition to it, Mahalanobis distance
may lead to convergence of the result in some cases. The time com- has been used in Ref. [121] to reduce the influence of the geometrical
plexity of the algorithm is given by O(knt), where k is number of shape of the different classes. The second limitation is dealt by de-
clusters, n is number of data points and t is number of iterations usually signing algorithms incorporating spatial information into account by
requiring k ≪ n and t ≪ n [109, 110]. The value of t depends on the many researchers as in Refs. [37, 47, 121-124].
initial starting cluster centres [110]. k-Means algorithm has few lim- Ahmed et al. redesigned the standard FCM algorithm incorporating
itations, say, the number of clusters k has to be initialized, the initial spatial contextual information, later called FCM_S [37] by using the
means also have to be chosen and it is affected by noise also. modified objective function to recoup intensity inhomogeneity and
Determining the actual number of clusters requires a-priori knowl- acknowledging the labels of pixels to be influenced by those of their
edge about the image. If the number of clusters chosen is the same as immediate neighbouring pixels. However, it computes the neighbour-
the real number of clusters, then k-means algorithm gives correct seg- hood term at each iteration which takes a lot of time. In order to
mentation; else, it ends up in giving faulty results. While clustering the overcome this problem, FCM_S is modified to FCM_S1 and FCM_S2 by
real data, the number of clusters is unknown beforehand, so the algo- introducing mean and median filtering (as proposed in Ref. [123]) that
rithm is run several times by performing clustering using different has also increased the robustness to outliers and noise. Computational
number of clusters which adds to computational cost and time. time was further reduced by Benaichouche et al. [121] using EnFCM by
Moreover, a lot of computations are required at every iteration to cal- working on gray level histogram rather than the image pixels.
culate the distance of each data point from its nearest cluster. Blending the ideas of FCM_S1, FCM_S2 and EnFCM, W. Cai
k-Means and hierarchical clustering are used to separate the posi- et al. [47] have proposed FGFCM clustering algorithm in which seg-
tion of tumorous tissues from the healthy ones in Ref. [111] by first menting time is significantly reduced by considering its dependency
rendering the original MRI image and then applying pseudo-color only on gray values rather than the image size. Moreover, a parameter α
transformation to convert it into RGB image to enhance its features. is used to balance between sensitivity to noise and preserving image
Juang and Wu [112] also works on the same idea and gave encouraging information, that demands expertise to be chosen, has been replaced by
results to determine exact lesion size and region. Ahmed and Mo- a new factor Sij, introduced as a local similarity measure that has re-
hamad [113] combined Perona and Malik [114] anisotropic diffusion duced uncertainty of correct segmentation. The method was suitable
model for image enhancement and k-means clustering for classifying when prior knowledge of noise does not exist, which generally happens.
different tissues and tumors and reliable results are witnessed. Furthermore, in spite of using the usual distance metric, a new dis-
Accurate determination of size and stage of tumor helps in taking similarity index influenced by local and non-local information is pro-
necessary and immediate steps to save the lives of tumor patients. posed by Wang et al. [125] which is proved to be better than above
Selvakumar et al. [115] have worked in this direction by using k-means algorithms but it is computationally expensive. Various variants of FCM
and Fuzzy C-Means algorithm. J. Vijay and J. Subhashini [38] proposed have been developed to make the standard FCM algorithm more robust
a method which gave efficient results in a comparatively lesser number to noise. Forouzanfar et al. [126] presented a neighbourhood attraction
of iterations. Madhukumar and Santhiyakumari [116] suggested k- dependent on the relative location and features of neighbouring pixels.
means, with histogram guided initialization has been proved to differ- FLICM algorithm that incorporates the local spatial information and
entiate all the six tissue classes (GM, WM, CSF, necrosis, enhancing gray level information in a fuzzy way was developed in Ref. [127]. Its
edema and background) on tumor edema complex MR images. In order performance is independent of the type of noise. The adaptive control
to tackle the drawback of initialization of cluster centres, Dha- factors introduced in objective function are achieved automatically
nachandra et al. [117] have implemented the use of subtractive clus- without any parameter selection like α which makes the segmentation
tering for initialization of k-means algorithm. It is accompanied by pre- process crisp. Song [128] improved segmentation accuracy by in-
processing step of partial stretching enhancement and post-processing corporating modified spatial constraints between the centre pixel and
step of median filtering to remove undesired regions, which finally neighbourhood pixels in similarity measurement.
results in improved segmentation quality. Numerous researchers have made an exclusive use of optimization
techniques to achieve better classification results. The researchers in
2.3.1.2. Fuzzy C-Means clustering. Ref. [121] proposed to initialize the FCM algorithm with the help of
FCM clustering algorithm was proposed by Bezdek in 1993 [118]. It PSO so that the system is not trapped in local minima. It helped to
works on the principle of assigning membership values uij to each choose the nearest initial solution to the global optimum. Menon and
feature data point xi,i = 1,2,…,n, xi ∈ℝd corresponding to the centres of Ramakrishnan [129] proposed the segmentation of MRI image using
the clusters cj, j = 1,2,…,k where k < n depending upon the Euclidean Artificial Bee Colony (ABC) with fitness function given in Ref. [130]
distance of the data point from the centre of the corresponding cluster along with FCM clustering. It not only segments the tumor but also
expressed by . The membership values satisfy the following provides with its intensity. A NIFCMGA is proposed in Ref. [131] which
properties:

254
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

not only helps to evaluate the optimal parameters for IFCM, but also preconvergence in case of ill-defined edges, oversegmentation in case of
reduces the effect of noise and outliers through neighbourhood mem- weak and diffused edges and noise, failure to segment heterogeneous
bership. Superiority in segmentation results of MRI image in presence of tumors and homogeneous tumors against the similar background. The
noise and IIH has been achieved in Ref. [124] by incorporating a comparisons with GVF, FVF and BVF proved the remarkable perfor-
conditional variable with each pixel which describes its significance in mance of CBAC.
generating the local membership values and the clusters. But, high IIH Many nature-inspired optimization techniques have been used to
levels cannot be dealt in with this approach. Incorporation of local overcome the limitations and solve the energy minimization problem.
contextual information and Gaussian function into the objective func- The control points in the traditional ACMs search in a small window for
tion in Ref. [55] has proved to be beneficial in IIH estimation and bias its new position to locate the edges which hinders accuracy in locating
field correction. concavities. A large search window is accompanied with more com-
putational time. In view to reduce this computational time, a MPSOS
2.3.2. Active contour models /deformable models has been proposed in Ref. [138]. Evolutionary algorithms followed by
ACM is a model-based segmentation technique, which is often called Greedy Algorithm have been used in Ref. [139] to automate the process
Deformable Models in its 3D version. It is mostly used in 3D image of selecting an optimal value for parameters to allow good detection of
segmentation [132]. A connected and continuous model is built for a edges. A MCSS has been introduced in Ref. [53] which is implemented
specific anatomic structure in model-based segmentation by in- with two geometries for search windows, RSSW and PSSW, out of which
corporating a-priori knowledge of the object such as shape, location, PSSW is preferred. MCSS was compared with traditional snakes and
and orientation. Given an initial propagating surface, deformable MPSOS. The assessments of the ACM strategies were performed through
models involve the evolution of the surface with the given speed similarity metrics; Jaccard coefficient, Dice coefficient, and Hausdorff
function such that it matches with the structure of the original object. distance.
Deformable models are capable of accommodating the significant
variability of biological structures over time and across different in- 2.3.2.2. Geometric deformable models.
dividuals. They are classified as parametric ACMs and geometric ACMs. Geometric models are based on curve evolution theory and level set
Parametric ACMs are implemented using techniques like finite element methods [40]. Instead of parameterizing a curve, it is depicted as a zero
method or splines whereas geometric models use the Eulerian frame- level set of a higher dimensional function whose evolution defines the
work like level sets [133]. initial value problem. Contour evolution is associated with speed
function of level sets. Here, the contour is independent of the curves
2.3.2.1. Parametric active contour models. parameterization [39]. So, they can handle topological changes.
Active contour models have been widely used for segmentation However, in case of tumor segmentation of MRI images wherein
since its inception. The parametric active contour model, also called images are generally noisy with ill-defined edges, geometric
snakes, was originally given by Kass, Witkin and Terzopoulois [22] deformable models may result in an inconsistent topology with
based on the gradient descent method. Active contours are simply respect to the actual object [140]. Caselles et. al. [24] proposed a
connected closed curves which track the boundary by matching the model for active contours based on a geometric partial differential
deformable model to the image curve, influenced by weighting forces equation called mean curvature motion equation. This scheme works
(external and internal forces), in order to minimize the defined energy well for objects having good contrast. When the object boundary is
functional. The energy function involves the parameters which control indistinct or has gaps, this contour tends to leak through the boundary.
the elasticity and rigidity of the curve respectively [24, 134]. External When boundaries of an object are not defined by gradient or with very
forces push the initial contour towards object contours and internal smooth boundaries, classical active contour models [22, 24, 27] are not
forces resist deformation. applicable. In the concerned direction, a method proposed by Tony F.
Though the model is simple and consistent, it has few limitations Chan [33] can be applied. In order to work well in case the boundaries
too. For instance, the snake has to be initialized close to the desired are ill-defined, a framework that uses edge-based information obtained
contour, else, it may find local minima away from the contour [133]. from gradients along with probability scores obtained from classifiers
The problem is solved by Cohen in Refs. [23, 25]. Dual active contour has been proposed in Ref. [141]. The method can acquire probability
models use two snakes to approach the desired contour from both inner scores from any classifier and can construct ESFs for any edge-based
and outer sides so as to improve the detection of global minima, active contour model using any level set method. Instead of using just
however low computational efficiency and complexity in its im- one ESF, it uses a group of ESFs which makes the process robust.
plementation restrict its use [2]. Traditional level set methods generally develop irregularities in the
Snakes did not perform well in presence of concavities and sharp LSF during its evolution which hampers its stability. In order to fix this
corners in the gray-level image which was overcome by the introduc- issue, re-initialization of LSF as and when required (when the LSF de-
tion of GVF as suggested by Xu and Prince [28, 30, 31]. However, GVF grades and requires reshaping for its evolution) has been proposed as a
requires prior object knowledge and appropriate parameter selection correcting measure. However, determining when LSF degrades and how
for good and reliable results. Moreover, it requires high computations. it should be corrected is itself of serious concern. In lieu of this, a dis-
Many researchers have introduced novel external force fields to over- tance regularization term has been introduced in potential function in
come the drawbacks of snakes. Sum et. al. proposed BVF [135] to deal Ref. [142] that maintains the shape of LSF. Moreover, the problem of
with problems of concavities and capture range. It used the interpola- poor convergence to boundaries has also been encountered which is
tion scheme to compute the potential values which helped reduce dealt in Ref. [134] by combining fuzzy clustering with deformable
computational burden. Wang et.al. proposed FVF in Ref. [136] to ex- models.
tract concave shaped tumors. It also addressed the issue of limited An automated 3D active contour method without edge has also been
capture range. It outperformed GVF and BVF based on the quantitative proposed in Ref. [9] that can identify tumor slices and segment them in
measures; mean, median and standard deviation. In Ref. [134], a robust volumetric MRI brain scans. Modify Level Set method combining geo-
method combining region based fuzzy clustering for getting an initial metric method with the statistical method of active contours has been
contour and deformable model for segmenting tumor region using GVF proposed in Ref. [1] to make initialization task automatic.
as external force field has been proposed that has solved the problems The region-based ACM methods, unlike edge-based ACMs, do not
associated with initialization and poor convergence to boundary con- depend on edges and gradient information of the image, hence more
cavities. Sachdeva et. al. [137] introduced CBAC that uses intensity as robust to noise. By incorporating the global image information, they
well as texture information in order to handle the problems like become robust to initialization. However, they assume the image

255
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

intensity to be homogeneous which does not often occur in the real misclassified pixels. Region-based ACM for segmentation and ANN
world. ACMs suffer from the limitation of handling IIH. Local image based LM algorithm for classification process are integrated together to
information is extracted through the introduction of a local binary fit- classify the MRI image efficiently as normal and tumorous as suggested
ting energy functional [143] and local image fitting energy func- by Shenbagarajan et al. [92]. k-NN and SVM are integrated with Chan-
tional [144] which helped in segmenting images with IIH. Wang et. al. Vese method [45] in order to get accurate segmentation results in lesser
took advantage of both global and local image information to frame a amount of time. Soltaninejad et al. [150] proposed an automated seg-
two-stage ACM [145]. In the first stage, a coarse segmentation contour mentation of brain tumor in mMRI images. In this study, a combination
is obtained by employing global energy fitting functional, which is then of FCNN with RF is presented. FCNN is used to form machine-learned
used as initial contour in second stage and local energy fitting func- features while RF for classification of normal and tumor tissues. The
tional is employed to finally extract the object boundaries. Bias cor- method was tested on MICCAI-BRATS 2017 challenge data sets. How-
rection is also an important method for addressing the IIH of MRI ever, the similar methodology for the desired objectives with improved
images. A modified model has been developed in Ref. [146] in which results is presented in Ref. [152]. More recently, an article presented in
clustering criterion energy function is defined by considering the dif- Ref. [153] demonstrates the state-of-the-art techniques participated in
ference between the measured image and estimated image in local re- BRATS challenge for the segmentation of brain tumors in mpMRI scans.
gion to obtain accurate segmentation results and an accurate estimation The paper concludes the competition by comparing various metho-
of the bias field. The problem of IIH is also dealt with introduction of dology and their findings. In this paper, the statement is made that the
locally statistical ACM [147]. Furthermore, incorporation of the mul- approaches based on machine learning are a good contender than
tiple image information into the singed pressure force term of SBGFRLS conventional methods at different levels of the competition. Finally, it is
model has made a significant contribution in solving blurred edge observed that on focusing the prediction of overall survival the Deep
problems in brain tumor segmentation [54]. The method effectively learning based techniques are superior to rest others.
segmented homogeneously enhanced, non-enhanced, heterogeneously- The hybrid approach presented Sharma et al. [93] to retrieve in-
enhanced and ring-enhanced tumor. A new LSM proposed by Kumar formation about the affected part and the size of a tumor from brain
et al. [148], based on probabilistic C-mean objective function and in- MRI images is based upon k-means and ANN. Fuzzy Inference System,
corporating lattice Boltzmann method is highly parallelizable, effective that uses GLCM for feature extraction is created, followed by thresh-
in tackling intensity inhomogeneity in image and robust to noise. It is olding, morphological operator and watershed segmentation for brain
computationally less complex. Region-based ACMs that use statistical tumor detection.
intensity information are sensitive to high mean intensity difference The combined approach of RFs and ACMs has been proposed in
between two consecutive regions. The difference in the mean intensities Ref. [107]. RFs are employed as feature learning kernels to learn
of the foreground and the background is quite significant which de- multiscale feature representations iteratively from multimodal MR vo-
creases the relevance of this method in MRI images. Therefore, Loca- lumes. Tumor structure is inferred from concatenated forests which is
lized ACM with background intensity compensation integrating an ad- taken as an initial contour to multiscale patch driven active contour
ditional automatic step to balance this difference is proposed in model for the final segmentation. This method is robust to tissues with
Ref. [149]. The introduction of HCSD enables the application of the low contrast. They require lesser computations and memory in contrast
approach on all sets of brain tumor images. It is compared to some to other machine learning methods like CNNs.
state-of-the-art methods for segmentation tasks and claimed to out-
perform them on grounds of computational time, dice coefficient, 3. Comparison and discussion
sensitivity, specificity, and the Hausdorff distance.
Table 2 shows an overview of recent methods based on different
2.4. Hybrid techniques parameters used for segmentation of brain tumor from brain MRI
images. Here, a comparative analysis in terms of the methodology used,
Hybrid techniques combine two or more techniques to provide types of abnormality and the type of supervision is presented. One may
segmentation results. It concatenates two or more methods by using observe here that most of the studies are focused towards high- and
their advantages and overcoming their disadvantages to get fruitful and low-grade gliomas. It is also analyzed through this comparison that
reliable outcomes. Parveen [151] suggested that FCM and SVM are increasing the size of the network will require more computational time
combined to acquire accurate classification of the diseased part of the and complexity.
brain. FCM is used for segmentation of doubtful, diseased region of the Tables 3 and 4 present a quantitative analysis of various state-of-
brain. Gray level run length matrix is used to extract features which are the-art methods through different evaluation parameters such as Dice,
used by SVM for classification. Precision, and Sensitivity for brain tumor segmentation of MRI images.
k-Means detects the tumor faster than FCM whereas FCM predicts Table 3 presents the results for two different cases of 110 and 274 for
tumor cells more accurately in comparison to k-means [43]. So, the BRATS 2015 data sets and Table 4 shows the comparative results for
proposed automatic algorithm in Ref. [43] uses the advantages of both both Challenge and Leaderboard data sets of BRATS 2013.
the methods to carry out efficient segmentation in lesser execution time However, very limited literature is available for comparing the re-
due to a reduced number of iterations. Tumor cluster is extracted and sults of BRATS 2015 data sets. Still, an effort is made to present the
then contoured using thresholding and active contours respectively as effectiveness of different methods for these data sets. Table 3 shows that
post-processing step. the performance of the method proposed by Kamnitsas et al. [12] is
A hybrid technique constituting LSVM classifier with Multi-Layer better for almost all the evaluation parameters than the other existing
perceptron based kernel functions and RBF kernel functions for classi- schemes. Here, it may be observed that the performance of Ref. [12] for
fication, and Fast Bounding Box for segmentation has been proposed in 110 cases is better when Ensemble with CRF is used while for 274 cases
Ref. [44]. Noise removal is done as a pre-processing step by median it is good for DeepMedic with CRF.
filtering followed by statistical based feature extraction using GLCM. Table 4 depicts that the performance of Zhao et al. [108] integrating
The classification accuracy is claimed to be 96.63% which is quite FCNN and CRF in a unified framework for Dice is better than others for
promising. Challenge data sets. However, the results of Tustison et al. [85] using
Machine learning and region-based ACMs have been integrated for the supervised learning capabilities of the RF model with regularized
effective segmentation using their respective advantages. Machine probabilistic segmentation and Kwon et al. [155] are good for sensi-
learning helps to deal with intensity inhomogeneities whereas region- tivity and Precision, respectively. For Leaderboard data sets, the per-
based ACMs deals efficiently with poorly defined boundaries and formance of Kwon et al. [155] using the combination of segmentation

256
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

and registration seems much better than the others. However, Zhao [21] Sujan M, Alam N, Abdullah S, Jahirul M. A segmentation based automated system
et al. [108] also provides the good results at some places for Leader- for brain tumor detection. Int J Comput Appl 2016;153(10):41–9.
[22] Kass M, Witkin A, Terzopoulos D. Snakes: active contour models. Int J Comput Vis
board data sets. As per authors' knowledge, no method is reported in the 1988;1(4):321–31.
literature that provides the best result for all the evaluation parameters. [23] Cohen L. On active contour models and balloons. CVGIP Image Underst
1991;53(2):211–8.
[24] Caselles V, Catt F, Coll B, Dibos F. A geometric model for active contours in image
4. Conclusions processing. Numer Math 1993;66(1):1–31.
[25] Cohen L D, Cohen I. Finite-element methods for active contour models and bal-
This paper presents an exhaustive study of various existing methods loons for 2-D and 3-D images. IEEE Trans Pattern Anal Mach Intell
1993;15(11):1131–47.
of brain tumor segmentation. The quantitative analysis through dif- [26] McInerney T, Terzopoulos D. A dynamic finite element surface model for seg-
ferent evaluation parameters among state-of-the-art methods helps the mentation and tracking in multidimensional medical images with application to
readers and clinicians not only to set the new directions to develop the cardiac 4D image analysis. Comput Med Imaging Graph 1995;19(1):69–83.
[27] Caselles V, Kimmel R, Sapiro G. Geodesic active contours. Int J Comput Vis
scope of research but also assists for accurate diagnosis of the tumor.
1997;22(1):61–79.
Finally, reviewing different methods, it has been observed that the [28] Xu C, Prince J L. Gradient vector flow: a new external force for snakes. Proc. IEEE
combination of CRF with FCNN and CRF with DeepMedic or Ensemble Computer Society Conf. Computer Vision and Pattern Recognition. 1997. p. 66–71.
are more effective towards fulfilling the requirements of segmentation [29] Gunn S R, Nixon M S. A robust snake implementation; a dual active contour. IEEE
Trans Pattern Anal Mach Intell 1997;19(1):63–8.
of brain tumor. [30] Xu C, Prince J L. Generalized gradient vector flow external forces for active con-
tours. Signal Process 1998;71(2):131–9.
Acknowledgments [31] Xu C, Prince J L. Snakes, shapes, and gradient vector flow. IEEE Trans Image
Process 1998;7(3):359–69.
[32] Giraldi G A, Strauss E, Oliveira A A F. A boundary extraction approach based on
The authors would like to thank the anonymous referees for their multi-resolution methods and the T-snakes framework. Proc. 13th Brazilian Symp.
valuable comments and suggestions that have greatly improved the Computer Graphics and Image Processing. 2000. p. 82–9.
[33] Chan T F, Vese L A. Active contours without edges. IEEE Trans Image Process
quality of this manuscript. 2001;10(2):266–77.
[34] Salman Y M. Modified technique for volumetric brain tumor measurements. J
References Biomed Sci Eng 2009;02:16–9.
[35] Sato M, Lakare S, Wan M, Kaufman A, Nakajima M. A gradient magnitude based
region growing algorithm for accurate segmentation. Proc. Int. Conf. Image
Processing. vol. 3. 2000. p. 448–51.
[36] Bajwa I, Asghar M, Naeem M. Learning-based improved seeded region growing
algorithm for brain tumor identification. Proceedings of the Pakistan Academy of
[1] Ben Rabeh A, Benzarti F, Amiri H. Segmentation of brain MRI using active contour
Sciences. vol. 54. 2017. p. 127–33.
model. Int J Imaging Syst Technol 2017;27(1):3–11.
[37] Ahmed M N, Yamany S M, Mohamed N, Farag A A, Moriarty T. A modified fuzzy C-
[2] Derraz F, Beladgham M, Khelif M. Application of active contour models in medical
means algorithm for bias field estimation and segmentation of MRI data. IEEE
image segmentation. Proc. Int. Conf. Information Technology: Coding and
Trans Med Imaging 2002;21(3):193–9.
Computing. vol. 2. 2004. p. 675–81.
[38] Vijay J, Subhashini J. An efficient brain tumor detection methodology using K-
[3] Meier R, Knecht U, Loosli T, Bauer S, Slotboom J, Wiest R. Clinical evaluation of a
means clustering algorithm. Int. Conf. Communication and Signal Processing.
fully-automatic segmentation method for longitudinal brain tumor volumetry. Sci
2013. p. 653–7.
Rep 2016;6:23376.
[39] Sethian J A. Numerical methods for propagating fronts. Variational Methods for
[4] Gordillo N, Montseny E, Sobrevilla P. State of the art survey on MRI brain tumor
Free Surface Interfaces. 1987.
segmentation. Magn Reson Imaging 2013;31(8):1426–38.
[40] Xu C, Yezzi A, Prince J L. On the relationship between parametric and geometric
[5] Bahadure N, Kumar Ray A, Pal Thethi H. Image analysis for MRI based brain tumor
active contours. 34th Asilomar Conf. Signals, Systems and Computers. vol. 1. 2000.
detection and feature extraction using biologically inspired BWT and SVM. Int J
p. 483–9.
Biomed Imaging 2017;2017:1–12.
[41] Ilunga-Mbuyamba E, Avina-Cervantes J G, Cepeda-Negrete J, Ibarra-Manzano M
[6] Wong K-P. Medical image segmentation: methods and applications in functional
A, Chalopin C. Automatic selection of localized region-based active contour
imaging. Topics in Biomedical Engineering International Book Series. Springer US;
models using image content analysis applied to brain tumor segmentation. Comput
2005. p. 111–82.
Biol Med 2017;91:69–79.
[7] Georgiadis P, Cavouras D, Kalatzis I, Daskalakis A, Kagadis G C, Sifaki K.
[42] El-Dahshan E-S A, Mohsen H M, Revett K, Salem A-B M. Computer-aided diagnosis
Improving brain tumor characterization on MRI by probabilistic neural networks
of human brain tumor through MRI: a survey and a new algorithm. Exp Syst Appl
and non-linear transformation of textural features. Comput Methods Prog Biomed
2014;41(11):5526–45.
2008;89(1):24–32.
[43] Abdel-Maksoud E, Elmogy M, Al-Awadi R. Brain tumor segmentation based on a
[8] Shen S, Sandham W, Granat M, Sterr A. MRI fuzzy segmentation of brain tissue
hybrid clustering technique. Egypt Inf J 2015;16(1):71–81.
using neighborhood attraction with neural-network optimization. IEEE Trans Inf
[44] Praveen G B, Agrawal A. Hybrid approach for brain tumor detection and classi-
Technol Biomed 2005;9(3):459–67.
fication in magnetic resonance images. Communication, Control and Intelligent
[9] Hasan A, Meziane F, Aspin R, Jalab H. Segmentation of brain tumors in MRI images
Systems (CCIS). 2015. p. 162–6.
using three-dimensional active contour without edge. Symmetry 2016;8(11):132.
[45] Pratondo A, Chui C-K, Ong S-H. Integrating machine learning with region-based
[10] Tonarelli L. Magnetic resonance imaging of brain tumor. CEwebsource.com.
active contour models in medical image segmentation. J. Vis. Commun. Image
[11] Hamid S Z, J P D, P W J, Tom M. Brain tumor segmentation and characterization
Represent. 2017;43:1–9.
by pattern analysis of multispectral NMR images. NMR Biomed 1998;11(4-
[46] Hung M-C, Wu J, Chang J-H, Yang D-L. An efficient k-means clustering algorithm
5):201–8.
using simple partitioning. J Inf Sci Eng 2005;21:1157–77.
[12] Kamnitsas K, Ledig C, Newcombe V F, Simpson J P, Kane A D, Menon D K. Efficient
[47] Cai W, Chen S, Zhang D. Fast and robust fuzzy c-means clustering algorithms in-
multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmen-
corporating local information for image segmentation. Pattern Recogn
tation. Med Image Anal 2017;36:61–78.
2007;40(3):825–38.
[13] Vaishnavee K B, Amshakala K. An automated MRI brain image segmentation and
[48] Zalik K R. An efficient k-means clustering algorithm. Pattern Recognition Letters
tumor detection using SOM-clustering and proximal support vector machine
2008;29(9):1385–91.
classifier. IEEE Int.Conf. Engineering and Technology (ICETECH). 2015.
[49] Moeskops P, Viergever M A, Mendrik A M, de Vries L S, Benders M J N L, Igum I.
[14] Tustison N J, Avants B B, Cook P A, Zheng Y, Egan A, Yushkevich P A. N4ITK:
Automatic segmentation of MR brain images with a convolutional neural network.
improved N3 bias correction. IEEE Trans Med Imaging 2010;29(6):1310–20.
IEEE Trans Med Imaging 2016;35(5):1252–61.
[15] Siddiqi K, Lauziere Y B, Tannenbaum A, Zucker S W. Area and length minimizing
[50] Soltaninejad M, Yang G, Lambrou T, Allinson N, Jones T L, Barrick T R. Automated
flows for shape segmentation. IEEE Trans Image Process. 1998;7(3):433–43.
brain tumour detection and segmentation using superpixel-based extremely ran-
[16] Sled J, Zijdenbos A, Evans A. A nonparametric method for automatic correction of
domized trees in FLAIR MRI. Int J Comput Assist Radiol Surg 2017;12(2):183–203.
intensity nonuniformity in MRI data. IEEE Trans Med Imaging 1998;17(1):87–97.
[51] Saouli R, Akil M, Kachouri R. Fully automatic brain tumor segmentation using end-
[17] Adams R, Bischof L. Seeded region growing. IEEE Trans Pattern Anal Mach Intell
to-end incremental deep neural networks in MRI images. Comput Methods Prog
1994;16(6):641–7.
Biomed 2018;166:39–49.
[18] Shih F Y, Cheng S. Automatic seeded region growing for color image segmentation.
[52] Soltaninejad M, Yang G, Lambrou T, Allinson N, Jones T L, Barrick T R. Supervised
Image Vis Comput 2005;23(10):877–86.
learning based multimodal MRI brain tumour segmentation using texture features
[19] Chaibou M, Conze P-H, Kalti K, Solaiman B, Mahjoub M. Adaptive strategy for
from supervoxels. Comput Methods Prog Biomed 2018;157:69–84.
superpixel-based region-growing image segmentation. J Electron Imaging
[53] Ilunga-Mbuyamba E, Cruz-Duarte J M, Avina-Cervantes J G, Correa-Cely C R,
2017;26.
Lindner D, Chalopin C. Active contours driven by cuckoo search strategy for brain
[20] Otsu N. A threshold selection method from gray-level histograms. IEEE Trans Syst
tumour images segmentation. Exp Syst Appl 2016;56:59–68.
Man Cybern 1979;9(1):62–6.
[54] Lok K H, Shi L, Zhu X, Wang D. Fast and robust brain tumor segmentation using

257
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

level set method with multiple image information. J X-Ray Sci Technol supervised brain tumor segmentation (simplified) with ANTsR. Neuroinformatics
2017;25(2):301–12. 2015;13(2):209–25.
[55] Mahata N, Kahali S, Adhikari S K, Sing J K. Local contextual information and [86] Pinto A, Pereira S, Correia H, Oliveira J, Rasteiro D M L D, da Silva C A B. Brain
gaussian function induced fuzzy clustering algorithm for brain MR image seg- tumour segmentation based on extremely randomized forest with high-level fea-
mentation and intensity inhomogeneity estimation. Appl Soft Comput tures. 37th Annual Int Conf IEEE Engineering in Medicine and Biology Society
2018;68:586–96. (EMBC). 2015. p. 3037–40.
[56] Raschke F, Barrick T R, Jones T L, Yang G, Ye X, Howe F A. Tissue-type mapping of [87] Chaddad A. Automated feature extraction in brain tumor by magnetic resonance
gliomas. NeuroImage Clin 2019;21:101648. imaging using Gaussian mixture models. J Biomed Imaging 2015;2015. 8:8–8:8.
[57] Gonzalez R C, Woods R E. Digital Image Processing. 2nd ed. Prentice [88] Goetz M, Weber C, Binczyk F, Polanska J, Tarnawski R, Bobek-Billewicz B. DALSA:
Hall0201180758; 2002. domain adaptation for supervised learning from sparsely annotated MR images.
[58] Aslam A, Khan E, Beg M S. Improved edge detection algorithm for brain tumor IEEE Trans Med Imaging 2016;35(1):184–96.
segmentation. Proc Comput Sci 2015;58:430–7. [89] Koley S, Sadhu A K, Mitra P, Chakraborty B, Chakraborty C. Delineation and di-
[59] Sujji G E, Lakshmi Y, Jiji G W. MRI brain image segmentation based on thresh- agnosis of brain tumors from post contrast T1-weighted MR images using rough
olding. Int J Adv Comput Res 2013;3(1):97. granular computing and random forest. Appl Soft Comput 2016;41:453–65.
[60] Ilhan U, Ilhan A. Brain tumor segmentation based on a new threshold approach. [90] Nayak D R, Dash R, Majhi B. Brain MR image classification using two-dimensional
Proc Comput Sci 2017;120:580–7. discrete wavelet transform and adaboost with random forests. Neurocomputing
[61] Taheri S, Ong S, Chong V. Level-set segmentation of brain tumors using a 2016;177:188–97.
threshold-based speed function. Image Vis Comput 2010;28(1):26–37. [91] Goetz M, Weber C, Bloecher J, Stieltjes B, Meinzer H-P, Maier-Hein K. Extremely
[62] Kim D Y, Park J W. Connectivity-based local adaptive thresholding for carotid Randomized Trees Based Brain Tumor Segmentation. 2014. p. 006–11.
artery segmentation using MRA images. Image Vis Comput 2005;23(14):1277–87. [92] Shenbagarajan A, Ramalingam V, Balasubramanian C, Palanivel S. Tumor diag-
[63] Lakare S, Kaufman A. 3D Segmentation Techniques for Medical Volumes. Center nosis in MRI brain image using ACM segmentation and ANN-LM classification
for Visual Computing, Department of Computer Science, State University of New techniques. Indian J. Sci. Technol. 2016;9(1).
York; 2000. p. 59–68. [93] Sharma M, Purohit G N, Mukherjee S. Information retrieves from brain MRI images
[64] Deng W, Xiao W, Deng H, Liu J. MRI brain tumor segmentation with region for tumor detection using hybrid technique K-means and artificial neural network
growing method based on the gradients and variances along and inside of the (KMANN). Networking Communication and Data Knowledge Engineering. 2018.
boundary curve. 3rd Int. Conf. Biomedical Engineering and Informatics. vol. 1. p. 145–57.
2010. p. 393–6. [94] Damodharan S, Raghavan D. Combining tissue segmentation and neural network
[65] Weglinski T, Fabijanska A. Brain tumor segmentation from MRI data sets using for brain tumor detection. Int Arab J Inf Technol 2015;12(1):42–52.
region growing approach. Proc. of 7th Int. Conf Perspective Technologies and [95] Wang S, Zhang Y, Dong Z, Du S, Ji G, Yan J. Feed-forward neural network opti-
Methods in MEMS Design (MEMSTECH). IEEE; 2011. p. 185–8. mized by hybridization of PSO and ABC for abnormal brain detection. Int J
[66] Kavitha A R, Chellamuthu C, Rupa K. An efficient approach for brain tumour Imaging Syst Technol 2015;25(2):153–64.
detection based on modified region growing and neural network in MRI images. [96] Muhammad N, Fazli W, Sajid A K. A simple and intelligent approach for brain MRI
Int. Conf. Computing, Electronics and Electrical Technologies (ICCEET). 2012. p. classification. J Intell Fuzzy Syst 2015;28(3):1127–35.
1087–95. [97] Wang S, Du S, Atangana A, Liu A, Lu Z. Application of stationary wavelet entropy
[67] Khalid N E A, Ibrahim S, Haniff P. MRI brain abnormalities segmentation using k- in pathological brain detection. Multimed Tools Appl 2018;77(3):3701–14.
nearest neighbors (k-NN). Int J Comput Sci Eng 2011;3(2):980–90. [98] Varuna Shree N, Kumar T N R. Identification and classification of brain tumor MRI
[68] Warfield S K, Kaus M, Jolesz F A, Kikinis R. Adaptive, template moderated, spa- images with feature extraction using DWT and probabilistic neural network. Brain
tially varying statistical classification. Med Image Anal 2000;4:43–55. Inf 2018;5(1):23–30.
[69] Steenwijk M D, Pouwels P J, Daams M, van Dalen J W, Caan M W, Richard E. [99] Pereira S, Pinto A, Alves V, Silva C A. Deep convolutional neural networks for the
Accurate white matter lesion segmentation by k nearest neighbor classification segmentation of gliomas in multi-sequence MRI. Brainlesion: Glioma, Multiple
with tissue type priors (kNN-TTPs). NeuroImage Clin 2013;3:462–9. Sclerosis, Stroke and Traumatic Brain Injuries. 2015. p. 131–43.
[70] Kumar T S, Rashmi K, Ramadoss S, Sandhya L K, Sangeetha T J. Brain tumor [100] Pereira S, Pinto A, Alves V, Silva C A. Brain tumor segmentation using convolu-
detection using SVM classifier. 3rd Int. Conf. Sensing, Signal Processing and tional neural networks in MRI images. IEEE Trans Med Imaging
Security (ICSSS). 2017. p. 318–23. 2016;35(5):1240–51.
[71] Ayachi R, Ben Amor N. Brain tumor segmentation using support vector machines. [101] Nyul L G, Udupa J K, Zhang X. New variants of a method of MRI scale standar-
European Conf. Symbolic and Quantitative Approaches to Reasoning and dization. IEEE Trans Med Imaging 2000;19(2):143–50.
Uncertainty. 2009. p. 736–47. [102] Zhang W, Li R, Deng H, Wang L, Lin W, Ji S. Deep convolutional neural networks
[72] Yang G, Jones T L, Howe F A, Barrick T R. Morphometric model for discrimination for multi-modality isointense infant brain image segmentation. NeuroImage
between glioblastoma multiforme and solitary metastasis using three-dimensional 2015;108:214–24.
shape analysis. Magn Reson Med 2016;75(6):2505–16. [103] Kleesiek J, Urban G, Hubert A, Schwarz D, Maier-Hein K, Bendszus M. Deep MRI
[73] Bauer S, Nolte L-P, Reyes M. Fully automatic segmentation of brain tumor images brain extraction: a 3D convolutional neural network for skull stripping.
using support vector machine classification in combination with hierarchical NeuroImage 2016;129:460–9.
conditional random field regularization. In: Fichtinger G, Martel A, Peters T, edi- [104] Dong H, Yang G, Liu F, Mo Y, Guo Y. Automatic brain tumor detection and seg-
tors. Medical Image Computing and Computer-Assisted Intervention — MICCAI. mentation using u-net based fully convolutional networks. Annual Conf. Medical
2011. p. 354–61. Image Understanding and Analysis. Springer; 2017. p. 506–17.
[74] Kharrat A, Gasmi K, Messaoud M B, Benamrane N, Abid M. A hybrid approach for [105] Havaei M, Davy A, Warde-Farley D, Biard A, Courville A, Bengio Y. Brain tumor
automatic classification of brain MRI using genetic algorithm and support vector segmentation with deep neural networks. Med Image Anal 2017;35:18–31.
machine. Leonardo J Sci 2010;17(1):71–82. [106] Shen D, Wu G, Suk H-I. Deep learning in medical image analysis. Ann Rev Biomed
[75] Zhang N, Ruan S, Lebonvallet S, Liao Q, Zhu Y. Kernel feature selection to fuse Eng 2017;19(1):221–48.
multi-spectral MRI images for brain tumor segmentation. Comput Vis Image [107] Ma C, Luo G, Wang K. Concatenated and connected random forests with multiscale
Underst 2011;115(2):256–69. patch driven active contour model for automated brain tumor segmentation of MR
[76] Wu W, Chen A Y C, Zhao L, Corso J J. Brain tumor detection and segmentation in a images. IEEE Trans Med Imaging 2018;37(8):1943–54.
CRF (conditional random fields) framework with pixel-pairwise affinity and su- [108] Zhao X, Wu Y, Song G, Li Z, Zhang Y, Fan Y. A deep learning model integrating
perpixel-level features. Int. J Comput Assist Radiol Surg 2014;9(2):241–53. FCNNs and CRFs for brain tumor segmentation. Med Image Anal 2018;43:98–111.
[77] Soltaninejad M, Ye X, Yang G, Allinson N, Lambrou T. Brain tumour grading in [109] Dong J, Qi M. K-means optimization algorithm for solving clustering problem.
different MRI protocols using SVM on statistical features. Med Image Underst Anal Second Int. Workshop on Knowledge Discovery and Data Mining. 2009. p. 52–5.
2014:259–64. [110] Na S, Xumin L, Yong G. Research on k-means clustering algorithm: an improved k-
[78] Soltaninejad M, Ye X, Yang G, Allinson N, Lambrou T. An image analysis approach means clustering algorithm. 3rd Int.Symp. Intelligent Information Technology and
to mri brain tumour grading. Oncol News 2015;9(6):204–7. Security Informatics. 2010. p. 63–7.
[79] Jones T L, Byrnes T J, Yang G, Howe F A, Bell B A, Barrick T R. Brain tumor [111] Wu M-N, Lin C-C, Chang C-C. Brain tumor detection using color-based K-means
classification using the diffusion tensor image segmentation (D-SEG) technique. clustering segmentation. iih-msp. 2007. p. 245–50.
Neuro-oncology 2014;17(3):466–76. [112] Juang L-H, Wu M-N. MRI brain lesion image detection based on color-converted K-
[80] Yang G, Jones T L, Barrick T R, Howe F A. Discrimination between glioblastoma means clustering segmentation. Measurement 2010;43(7):941–9.
multiforme and solitary metastasis using morphological features derived from the [113] Ahmed M M, Mohamad D B. Segmentation of brain MR images for tumor ex-
p: q tensor decomposition of diffusion tensor imaging. NMR Biomed traction by combining K-means clustering and Perona-Malik anisotropic diffusion
2014;27(9):1103–11. model. Int J Image Process 2008;2(1):27–34.
[81] Mangasarian O L, Wild E W. Proximal support vector machine classifiers. Proc. [114] Perona P, Malik J. Scale-space and edge detection using anisotropic diffusion. IEEE
KDD-2001: Knowledge Discovery and Data Mining. 2001. p. 77–86. Trans Pattern Anal Mach Intell 1990;12(7):629–39.
[82] Jang J S R. ANFIS: Adaptive-network-based fuzzy inference system. IEEE Trans [115] Selvakumar J, Lakshmi A, Arivoli T. Brain tumor segmentation and its area cal-
Syst Man Cybern 1993;23(3):665–85. culation in brain MR images using K-mean clustering and fuzzy C-mean algorithm.
[83] Lahmiri S. Glioma detection based on multi-fractal features of segmented brain Int. Conf. Advances in Engineering, Science and Management (ICAESM). 2012. p.
MRI by particle swarm optimization techniques. Biomed Signal Process Control 186–90.
2017;31:148–55. [116] Madhukumar S, Santhiyakumari N. Evaluation of k-means and fuzzy C-means
[84] Breiman L. Random forests. Mach Learn 2001;45(1):5–32. segmentation on MR images of brain. Egypt J Radiol Nucl Med 2015;46(2):475–9.
[85] Tustison N J, Shrinidhi K L, Wintermark M, Durst C R, Kandel B M, Gee J C. [117] Dhanachandra N, Manglem K, Chanu Y J. Image segmentation using k-means
Optimal symmetric multimodal templates and concatenated random forests for clustering algorithm and subtractive clustering algorithm. Procedia Computer

258
A. Wadhwa, et al. Magnetic Resonance Imaging 61 (2019) 247–259

Science 2015;54:764–71. particle swarm optimization. Expert Syst Appl 2009;36(3, Part 1):5348–52.
[118] Bezdek J C, Hall L O, Clarke L P. Review of MR image segmentation techniques [139] Zohra B F, Nacra B, Abdelmalik T A. Adjustment of active contour parameters in
using pattern recognition. Med Phys 1993;20(4):1033–48. brain MRI segmentation using evolution strategies. 4th Int. Conf. on Electrical
[119] Jajuga K. L1-norm based fuzzy clustering. Fuzzy Sets Syst 1991;39(1):43–50. Engineering (ICEE). 2015. p. 1–7.
[120] Hathaway R J, Bezdek J C, Hu Y. Generalized fuzzy c-means clustering strategies [140] Xu C, Pham D L, Prince J L. Image segmentation using deformable models. Handb
using Lp norm distances. Trans Fuzzy Syst 2000;8(5):576–82. Med Imaging 2000;2:129–74.
[121] Benaichouche A, Oulhadj H, Siarry P. Improved spatial fuzzy c-means clustering [141] Pratondo A, Chui C, Ong S. Robust edge-stop functions for edge-based active
for image segmentation using PSO initialization, Mahalanobis distance and post- contour models in medical image segmentation. IEEE Signal Process Lett
segmentation correction. Dig Signal Process 2013;23(5):1390–400. 2016;23(2):222–6.
[122] Szilagyi L, Benyo Z, Szilagyi S M, Adam H S. MR brain image segmentation using [142] Li C, Xu C, Gui C, Fox M D. Distance regularized level set evolution and its ap-
an enhanced fuzzy C-means algorithm. Proc. 25th Annual Int. Conf. Engineering in plication to image segmentation. IEEE Trans Image Process 2010;19(12):3243.
Medicine and Biology Society. vol. 1. 2003. p. 724–6. [143] Li C, Kao C, Gore J C, Ding Z. Implicit active contours driven by local binary fitting
[123] Chen S, Zhang D. Robust image segmentation using FCM with spatial constraints energy. IEEE Conf. Computer Vision and Pattern Recognition. 2007. p. 1–7.
based on new kernel-induced distance measure. IEEE Trans Syst Man Cybern B [144] Zhang K, Song H, Zhang L. Active contours driven by local image fitting energy.
(Cybern) 2004;34(4):1907–16. Pattern Recogn 2010;43(4):1199–206.
[124] Adhikari S K, Sing J K, Basu D K, Nasipuri M. Conditional spatial fuzzy C-means [145] Wang H, Huang T-Z, Xu Z, Wang Y. A two-stage image segmentation via global
clustering algorithm for segmentation of MRI images. Appl Soft Comput and local region active contours. Neurocomputing 2016;205:130–40.
2015;34:758–69. [146] Huang C, Zeng L. An active contour model for the segmentation of images with
[125] Wang J, Kong J, Lu Y, Qi M, Zhang B. A modified FCM algorithm for MRI brain intensity inhomogeneities and bias field estimation. PloS One. 2015.
image segmentation using both local and non-local spatial constraints. Comput [147] Mostaar A, Houshyari M, Badieyan S. A novel active contour model for MRI brain
Med Imaging Graph 2008;32(8):685–98. segmentation used in radiotherapy treatment planning. Electron Phys
[126] Forouzanfar M, Forghani N, Teshnehlab M. Parameter optimization of improved 2016;8(5):2443–51.
fuzzy C-means clustering algorithm for brain MR image segmentation. Eng Appl [148] Kumar R, Rani S, Sarkar A, Talukdar F A. GPU-based Level Set Method for MRI
Artif Intell 2010;23(2). Brain Tumor Segmentation Using Modified Probabilistic Clustering. IGI Global;
[127] Krinidis S, Chatzis V. A robust fuzzy local information C-means clustering algo- 2017. p. 1053–78.
rithm. IEEE Trans Image Process 2010;19(5):1328–37. [149] Ilunga-Mbuyamba E, Avina-Cervantes J G, Garcia-Perez A, de Jesus Romero-
[128] Song J-H, Cong W, Li J. A fuzzy c-means clustering algorithm for image seg- Troncoso R, Aguirre-Ramos H, Cruz-Aceves I. Localized active contour model with
mentation using nonlinear weighted local information. J Inf Hiding Multimedia background intensity compensation applied on automatic MR brain tumor seg-
Signal Process 2017;8(9):1–11. mentation. Neurocomputing 2017;220:84–97.
[129] Menon N, Ramakrishnan R. Brain tumor segmentation in MRI images using un- [150] Soltaninejad M, Zhang L, Lambrou T, Yang G, Allinson N, Ye X. MRI brain tumor
supervised artificial bee colony algorithm and FCM clustering. Int. Conf. segmentation and patient survival prediction using random forests and fully
Communications and Signal Processing (ICCSP). 2015. p. 0006–9. convolutional networks. Int. MICCAI Brainlesion Workshop. Springer; 2017. p.
[130] Ma M, Liang J, Guo M, Fan Y, Yin Y. SAR image segmentation based on artificial 204–15.
bee colony algorithm. Appl Soft Comput 2011;11(8):5205–14. [151] Parveen. Singh A. Detection of brain tumor in MRI images, using combination of
[131] Huang C-W, Lin K-P, Wu M-C, Hung K-C, Liu G-S, Jen C-H. Intuitionistic fuzzy C- fuzzy C-means and SVM. 2nd Int. Conf. Signal Processing and Integrated Networks
means clustering algorithm with neighborhood attraction in segmenting medical (SPIN). 2015. p. 98–102.
image. Soft Comput 2015;19(2):459–70. [152] Soltaninejad M, Zhang L, Lambrou T, Yang G, Allinson N, Ye X. MRI brain tumor
[132] Chang H-H, Valentino D J. An electrostatic deformable model for medical image segmentation using random forests and fully convolutional networks.
segmentation. Comput Med Imaging Graph 2008;32(1):22–35. International MICCAI Brainlesion Workshop. 2017. p. 279–83.
[133] Srikrishnan V, Chaudhuri S, Roy S D, Sevcovic D. On stabilisation of parametric [153] Bakas S, Reyes M, Jakab A, Bauer S, Rempfler M, Crimi A. Identifying the best
active contours. IEEE Conf. Computer Vision and Pattern Recognition, CVPR’07. machine learning algorithms for brain tumor segmentation, progression assess-
2007. p. 1–6. ment, and overall survival prediction in the brats challenge arXiv preprint
[134] Rajendran A, Dhanasekaran R. Fuzzy clustering and deformable model for tumor arXiv:181102629; 2018.
segmentation on MRI brain image: a combined approach. Proc Eng [154] Bakas S, Zeng K, Sotiras A, Rathore S, Akbari H, Gaonkar B. GLISTRboost: com-
2012;30:327–33. bining multimodal MRI segmentation, registration, and biophysical tumor growth
[135] Sum K, Cheung P Y. Boundary vector field for parametric active contours. Pattern modeling with gradient boosting machines for glioma segmentation. Int.
Recogn 2007;40(6):1635–45. Workshop on Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain
[136] Wang T, Cheng I, Basu* A. Fluid vector flow and applications in brain tumor Injuries. Springer; 2015. p. 144–55.
segmentation. IEEE Tran Biomed Eng 2009;56(3):781–9. [155] Kwon D, Shinohara R T, Akbari H, Davatzikos C. Combining generative models for
[137] Sachdeva J, Kumar V, Gupta I, Khandelwal N, Ahuja C K. A novel content-based multifocal glioma segmentation and registration. Int. Conf. Medical Image
active contour model for brain tumor segmentation. Magn Reson Imaging Computing and Computer-assisted Intervention. Springer; 2014. p. 763–70.
2012;30(5):694–715. [156] Davy A, Havaei M, Warder-Farley D, Biard A, Tran L, Jodoin P. Brain tumor
[138] Tseng C-C, Hsieh J-G, Jeng J-H. Active contour model via multi-population segmentation with deep neural networks. Proc. of BRATS-MICCAI. 2014.

259