The Visual Computer (2022) 38:107–117

https://doi.org/10.1007/s00371-020-02005-1

ORIGINAL ARTICLE

Brain tumor classification based on hybrid approach

Wadhah Ayadi1 · Imen Charfi2 · Wajdi Elhamzi3 · Mohamed Atri4

Accepted: 17 October 2020 / Published online: 4 November 2020

© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
Various computer systems have attracted more researchers’ attention to arrive at a qualitative diagnosis in a few times.
Different brain tumor classification approaches are proposed due to lesion complexity. This complexity makes the early tumor
diagnosis using magnetic resonance images (MRI) a hard step. However, the accuracy of these techniques requires a significant
amelioration to meet the needs of real-world diagnostic situations. We aim to classify three brain tumor types in this paper. A
new technique is suggested which provides excellent results and surpasses the previous schemes. The proposed scheme makes
use of the normalization, dense speeded up robust features, and histogram of gradient approaches to ameliorate MRI quality
and generate a discriminative feature set. We exploit support vector machine in the classification step. The suggested system
is benchmarked on an important dataset. The accuracy achieved based on this scheme is 90.27%. This method surpassed
the most recent system according to experimental results. The results were earned through a strict statistical analysis (k-fold
cross-validation), which proves the reliability and robustness of the suggested method.

Keywords MRI · Classification · Brain tumor · DSURF · HoG

1 Introduction that significantly hold back to patient’s activities [1]. Brain

tumor is considered through the main causes of the increase in
The brain is very complex and sensitive, and it controlled the mortality in the whole world, which augment the burden for
overall body functionality. It can be affected by a tumor. The hospitals and society. Recently, many comprehensive statisti-
tumor can make changes in brain behavior and structure. So, cal studies concerning brain cancer in the world are published
any brain damage affects the body badly. [2–4].
Brain tumor represents a group of abnormal cells where The detected cases of brain cancer in the USA are
they are produced from parenchyma in the brain or in neigh- increased from 1.5 million to 1.658 million between 2013
boring parts. Brain cancer can conduct to extreme disabilities and 2015 [1]. In 2018, the approximated number of deaths
is near 16,830 [5]. In the aim to generate good brain image
B Wajdi Elhamzi without using radioactive traces or ionizing radiation, MRI
w.elhamzi@essths.rnu.tn
is exploited. It provides huge information around the brain
Wadhah Ayadi tissues that are helpful for clinical diagnosis and medical
wadhah.ayadi@fsm.rnu.tn
studies. In comparison with various techniques, the MRI gen-
Imen Charfi erates several tissues contrasts, which make it the favorite
imenecharfi@yahoo.fr
method for brain tumor treatment and diagnosis [1].
Mohamed Atri Recently, the analysis of brain magnetic images receives
matri@kku.edu.sa
large attention. However, brain pathology is not an ordi-
1 Laboratory of Electronics and Microelectronics, University of nary work. The classical scheme is by human inspection
Monastir, Monastir 5000, Tunisia that strongly relies on radiologist’s experiences. A huge data
2 Higher School of Sciences and Technology of Hammam can generate difficulty for radiologist to examine manually
Sousse, University of Sousse, Sousse, Tunisia the brain images. Besides, the manual methods existed are
3 College of Computer Sciences and Engineering, Prince tedious and expensive and need time. Therefore, divers sys-
Sattam Bin Abdulaziz University, Al-Kharj, Saudi Arabia tems were suggested to assist the radiologist to make a
4 College of Computer Science, King Khalid University, Abha, decision [6–14]. These systems can take a significant role
Saudi Arabia

123
108 W. Ayadi et al.

to facilitate radiologist’s diagnosis, prognosis, pre-surgical classification outperforms other state-of-the-art methods and
and postsurgical processes, etc. [1, 15]. provide 93.7% as accuracy.
Many attempts have proposed various semiautomatic or Hang et al. [24] suggested a new method for cancer
automatic techniques for brain tumor detection [16–19]. In diagnosis based on gene expression data based on sparse rep-
light of this, a new automatic approach is suggested, which resentations. The proposed scheme is applied to six datasets
categorizes brain tumors into several kinds. It represents a and the results are compared with previous methods. The
challenge compared to traditional methods. Our contribution experimental results have shown that the performance of the
comprises the following key points: proposed technique is more efficient compared to the previ-
ous technique.
• New classification scheme is suggested in aim to classify In order to help the radiologists in MRI classification,
brain tumors (three types). Sachdeva et al. [14] have suggested a semiautomatic clas-
• The impact of using complete MR images is explored, sification scheme that contains varied steps. To detect tumor
where it achieves ameliorated results in multi-class classi- areas, a content-based active contour system that allows the
fication. radiologist to manually indicate the region of interest (ROI),
• The suggested scheme classifies brain tumors into multi- which is saved as a segmented ROI (SROI) is applied in the
class, compared to the traditional methods, which rely on first step. Then, 71 texture and intensity features are extracted
expensive binary classification. using the SROI. Optimal feature selection is performed with
the application of Genetic Algorithm (GA). The last phase
This work is divided into eight sections. Some related classifies the chosen features using two classifier SVM and
works are briefly given in Sect. 2. Section 3 is designated to ANN. The suggested scheme is tested on different datasets.
explain the proposed scheme. The measurements are outlined The first dataset contains 428 MR images and the second
in Sect. 4. Section 5 is provided to introduce the dataset used. includes 260. The first set of images contains six tumors
Section 6 presents the results and Section 7 discussion, and categories such as glioblastoma multiforme, meningioma,
Sect. 8 concludes. astrocytoma, childhood tumor-medulloblastoma, and sec-
ondary tumor-metastatic. The second dataset contains only
three tumor categories, which are AS, MEN, and low-grade
2 Related works glioma. The suggested GA-SVM aims to find a preliminary
probability for tumor category, while GA-ANN aims to con-
It is important to extract image features and classify it in var- firm the accuracy. The performance calculated on the first
ious classes. Generally, the features can be grouped as basic group of images shows that GA-based approach enhanced
or complex [20, 21]. The basic features can include contrast, SVM accuracy to 91.7% while ANN accuracy to 94.9%.
homogeneity, sum entropy, correlation, average, variance, SVM accuracy has raised to 89% and 94.1% for ANN in
entropy, kurtosis, etc. Several complex features are calcu- the second group of images. The results demonstrate that
lated using the basic features [22]. the classifier GA-ANN offered the highest results compared
In the literature, various studies are reported that are to GA-SVM. Besides, the GA-SVM yields the speed, while
dedicated to brain tumor classification. However, multiple GA-ANN yields the accuracy. According to results, the sug-
works are exploited for a binary classification, while a few gested scheme has acceptable performance and can assist
works are founded in multi-class classification. The clas- radiologists in taking a better decision to classify brain
sification approaches are based either on segmentation or tumors.
non-segmentation for the tumor area. We aim in this section Other work is interested in classifying meningioma,
to discuss some approaches. pituitary tumor, and glioma [26]. New capsule network archi-
Several researchers exploit the sparse representation for tecture (CapsNets) for brain tumor classification is proposed
image classification such as [23–25]. to maximize the accuracy result. The suggested architecture
Al-Shaikhli et al. [23] propose a novel approach to multi- extracts different features from segmented images as edges
class brain tumor classification based on sparse coding and and corners and it is tested on a public database, which con-
dictionary learning. This suggested approach exploits topo- tains 3064 MRI images. The segmented and the whole brain
logical and texture features in order to build and learn a images are provided in this dataset. The new architecture con-
dictionary. Three medical datasets are exploited to collect tains convolutional layers with 64 feature maps. The accuracy
50 normal brain cases, 50 cases with brain glioma, 50 cases achieved for proposed CapsNets is about 78% using the
with brain glioblastoma, and 50 cases with brain metastatic whole image and 86.56% using the tumor-segmented region.
carcinoma. In the testing phase, a tenfold cross-validation is Other convolution neural networks (CNN) architecture is
used to evaluate the performance of the classification. Exper- proposed. It contains various convolution and fully connected
imental results demonstrate that the sparse coding-based (FC) layers. The results achieved are about 61.97% for the

123
Brain tumor classification based on hybrid approach 109

whole image and 72.13% for the tumor-segmented region. niques are exploited to minimize the computation and extract
According to results, the CapsNet surmount CNN for the the important features as LDA and PCA. The obtained fea-
two types of images. tures are exploited as input to the KNN and SVM classifiers.
A new method is proposed by Liu et al. [27] in the aim The dataset is taken from Harvard Medical School. They con-
to distinguish among primary central nervous system lym- tain glioma, Alzheimer, meningioma, Alzheimer plus visual
phoma (PCNSL) and glioblastoma multiform (GBM). Gabor agnosia, pick, huntington, sarcoma, and normal. For each
wavelet is exploited to detect features from T1 weighted class, ten images are exploited, i.e., a total of 80 images. All
images. Various frequencies and directions are exploited those images have a size equal to 256 × 256 pixels and T2-
on a Gabor wavelet transformation to extract discriminant weighted in the axial plane. The gained accuracy is about
features. The wavelet is utilized to minimize noise in the pre- 98%.
processing phase. SVM classifier is exploited to segment the Caulo et al. [30] suggested a new scheme to classify a
tumor area in the aim to evaluate Gabor texture features. Lin- glioma grade. Three schemes are suggested and evaluated
ear discriminant analysis (LDA) is exploited as a last step. based on 110 images patients with the extraction of histolog-
The result obtained proved that the suggested method can ical features. The highest accuracy achieved is about 95.5%,
distinguish various categories of tumor images with a higher 83.9% of sensitivity and 96.2% of specificity. The use of multi
accuracy rate. parametric sequences generates more information useful in
In order to classify three brain tumor categories (glioma, the classification process.
meningioma, and pituitary tumor), a new scheme is pro- A new technique was proposed by Lin et al. [31] to
posed by Cheng et al. [28]. They use images in T1 modality classify meningioma tumor in a different grade. Grade I
with contrast-enhanced. The public dataset contains 3064 contains the noncancerous, which are slow-growing tumors.
slices from 233 patients acquired from 2005 to 2010 from Grade II contains cancerous and noncancerous tumors. The
China Hospitals. It includes 708 meningiomas, 1426 gliomas, grade III contains cancerous tumors, which can grow quickly.
and 930 pituitary tumors. The augmented ROIs with manual Different features are exploited as contextual and radiolog-
segmentation are used instead of using original ROIs since ical features. No segmentation or preprocessing process is
tumor surrounding tissues can offer such indicative features. performed. In classification step, the authors used multiple
Intensity normalization is realized before segmentation to logistic regression. The proposed scheme is tested using 120
enhance image quality. Numerous features are extracted as private patients MRI images, were 90 with Grade I and 30
intensity histogram, BoW, and GLCM. The extracted features with Grade II or III. They exploited several sequences as
are reduced using LDA. Finally, three classifiers are utilized FLAIR T1 and T2. DWI transformation is utilized to extract
to label the feature, which are SVM, sparse representation- features. The results are acceptable with the exploited dataset.
based classification and K-Nearest Neighbor (KNN). A huge However, this kind of method requires many large dataset to
data size is exploited, and among all the feature technique ensure its validity.
used, the BoW produces the highest results that achieve Deepak et al. [32] suggested a new technique for MRI
91.28%. This technique is effective for the classification brain tumor classification. They exploit the deep transfer
according to the provided results. A manual segmentation learning and the pre-trained GoogLeNet with the aim to
approach is used and such parameters are tuned manually, extract MRI features. Figshare dataset is used to test the
which can restrict the scalability of the suggested system. performance of the proposed method based on fivefold
Other works do not apply any segmentation step for MRI cross-validation. The classification accuracy achieved by this
brain tumor classification. Hemanth et al. [29] used modified model is about 92.3%. This result was further improved to
Neural Network to analyze abnormal brain images. 540 pri- 97.8% and 98% with the use of SVM and KNN, respectively,
vate MR brain images were exploited to test the suggested as a CNN feature classifier. Despite the fewer training sam-
method. The dataset consists of four tumors class which are ples, the suggested technique outperforms all state-of-the-art
namely astrocytoma, meningioma, metastase, and glioma. methods.
The used images are of 256*256 size. A normalization is per- CNN is exploited also by Ahmad et al. [33] with the aim
formed as a preprocessing step. Eight features are acquired to classify MRI brain tumor images. The proposed technique
based on the first-order histogram and GLCM. The suggested uses Discrete Wavelet Transform (DWT) and a CNN model.
method provides promising results that can reach 95% as sen- 510 MRI images are selected from the Figshare dataset in
sitivity, 98% as specificity and 98% as accuracy. order to test the model performance. The experimental results
Kalbkhani et al. [12] used the generalized autoregressive achieve 99.3% as overall accuracy.
conditional heteroscedasticity model to distinguish seven Paul et al. [34] applied three different classifiers: CNN,
brain diseases with normal. With this approach, DWT is cal- fully connected neural network, and random forest in the
culated. The model parameters define the feature vector. A aim to classify MRI brain tumor. CNN provides the highest
normalization process is applied to this vector. Several tech- accuracy rate, which attained 90.26%. The proposed model

123
110 W. Ayadi et al.

contains various layers as convolutional, MaxPool, and fully method choice is not randomly but through the good experi-
connected. mental results compared to some basic filters as the median.
Abiwinanda et al. [35] investigated the application of CNN The intensity normalization is performed using Eq. (1):
for this dataset and designed seven various neural networks.
The second model provided the highest performance, which MaxNew − MinNew
ImgNew  (Img − Min) + MinNew (1)
contains two convolutional and one fully connected layers. Max − Min
This simple model and without any previous segmentation
achieve 98.51% as training accuracy and 84.19% as test accu- where Img and ImgNew are the original and the newest image
racy. successively. The Min and Max represent the low and high
Different preprocessing manners are investigated by Tahir intensity for Img, while MinNew and MaxNew are correspon-
et al. [36] to ameliorate the classification result. They grouped dent to ImgNew .
these techniques into three groups: noise removal, edge
detection, and contrast enhancement. The possible combina- 3.2 Feature extraction
tions are applied to different image sets. The authors affirm
that the combination of various preprocessing techniques is 3.2.1 Dense feature points
more beneficial than applying a single technique. The SVM
classifier was exploited and reported 86% as the highest accu- The DSURF descriptor [38] is a variant of Speeded Up
racy on Figshare dataset. Robust Features (SURF). SURF represents a feature point
Various schemes were presented in the literature for the detector. It is exploited also as a feature point descriptor [39].
classification of brain tumors. It is very essential to analyze Firstly and after key-point identifying, the process defines an
MRI brain images before any surgical resection and it is orientation around this key-point in a circular region. The
important to extract as much information as possible to avoid region is then aligned to an orientation in the aim to extract
pain and time. Different schemes are exploited to extract the SURF descriptor, which can be 64 or 128 dimensions. The
important feature. Most researchers used small datasets to major difference among SURF and DSURF is the technique
examine their proposed approaches. However, we note that of key-point selection. Dense feature points were always sit-
with the increase in dataset size, the accuracy can be declined. uated closely together along a grid with a particular step size.
Several researchers have used various datasets, which limits A feature descriptor was calculated for each key-point. A sig-
the comparison of the performance. A new scheme is sug- nificant feature was gained in this way [40]. Due to the huge
gested to exceed these drawbacks. feature extracted from the image, DSURF is more effective
than SURF when foreknowledge is limited.

3 Suggested method 3.2.2 HoG

New scheme was proposed in the aim to generate an accurate Due to its success in complex problem, HOG has attracted
classification system. The suggested approach is illustrated attention as feature extractor and it has been much exploited
in Fig. 1. Our technique comprises three steps, which are in the literature [41]. HOG feature is generally exploited in
image normalization to enhance image quality, and then we different fields like pedestrian detection [42], object detection
applied DSURF and HoG to extract the important feature. [43, 44], image registration [45], medical image classification
Finally, SVM was exploited as a classifier in the last step. [46, 47], etc. It estimates the number of times the oriented
gradient occurs in a specific area in the image. It identifies
3.1 Intensity normalization the edge information that is useful in classification. With
HOG method, the image was partitioned into connected and
In this study, a hybrid method (DSURF + HoG) is used to small regions named cells. The orientations of the edge or
extract feature. The MRI intensity values are sensitive to the HOG directions for every cell are calculated. The obtained
acquisition conditions. However, it has not a fixed mean- histograms are combined to create the descriptor [48–50].
ing and it varies considerably within or between subjects. The information gained from different regions in an image
Furthermore, owing the intensity variations, all images are is presented as a feature vector. In HOG process, all blocks
normalized through the min–max technique. This normal- create the density of its intensity gradients. The gradients,
ization considered necessary in texture analysis and it is a ∂ f (x, y)/∂ x and ∂ f (x, y)/∂ y, are calculated using, respec-
needed step to preprocess these images [37]. The process tively, Eqs. (2) and (3):
calculates the high (Max) and low (Min) intensity of the
brain image. The image is mapped with new intensity val- ∂ f (x, y) f (x + 1, y) − f (x − 1, y)
ues (MaxNew ,MinNew ) using the min–max technique. This Gx   (2)
∂x (x + 1) − (x − 1)

123
Brain tumor classification based on hybrid approach 111

Fig. 1 The suggested approach

∂ f (x, y) f (x, y + 1) − f (x, y − 1) TP

Gy   (3) Sensitivity (Recall)  (6)
∂y (y + 1) − (y − 1) TP + FN
TP
The use of HOG descriptor represents an efficient man- Precision  (7)
TP + FP
ner to detect textures and deformable objects. It has strong
robustness of the brightness variation. Moreover, it represents Recall ∗ Precision
F1 Score  2 ∗ (8)
a suitable way also to detect abnormal tissues [51]. Recall + Precision

where TP identifies the True Positive, TN identifies the True

3.3 SVM Negative, FP identifies the False Positive and FN identifies
the False Negative.
SVM is suggested by Cortes and Vapnik [52] for the binary
classification, which lacks a statistical sample. It is applied to
various problems as face recognition [53], text categorization 5 Image data
[54], cancer diagnosis [55], and classification [56–58]. This
classifier can exploit several kernels functions like linear, In this work, the dataset exploited was collected from Nan-
radial basis function, polynomial, etc. It goals to differentiate fang Hospital and General Hospital, China, from 2005 to
the input to a decision surface, which represents the hyper- 2010. The collected data contains 233 patients suffering
plane. Its principal role is to maximize the margins through from brain tumors, wherein three classes of tumors are
two classes of the hyperplane [59, 60]. In general, the prob- apparent, namely meningioma, pituitary, and glioma tumors.
lems of real world are not linear. In this situation, a nonlinear The images have a size of 512 × 512 pixels. The dataset
classification is desired. SVM can also divide the classes in comprises 3064 slices, including 708 meningioma, 1426
a nonlinear fashion. In this situation, the data will be pro- glioma, and 930 pituitary tumors [61]. Table 1 represents
jected to a higher dimension to make the problem linearly the dataset details exploited in our study. For every patient,
separable. three radiologists examined independently the MRI image
to define the pathological type. This dataset is complicated,
and some tumors are similar in color, shape, position,etc.
4 Performance evaluation Some examples of MRI images are given in Fig. 2. Fig-
ures 2a and 3b present two patients having the same category
The efficiency of image classification into three classes is of tumors, where they exhibit different appearances. Con-
measured through some performance measures such as accu- versely, a visual similarity is shown in Figs. 2c and 3d, where
racy, specificity, sensitivity, precision, and F1-score. For are containing different pathological categories. In this work,
each category, the average measures are estimated to define two images include the same tumor class are considered as
classifier performance. The formula for calculating these per- similar, else, they are considered dissimilar.
formances are written as below in Eqs. 4, 5, 6, 7 and 8,
respectively:
6 Experimental results
TP + TN
Accuracy  (4)
(TP + TN + FP + FN) This section verifies the suggested scheme validity using
TN brain MRI images, which contained different classes of
Specificity  (5) tumors. The new scheme is tested using cross-validation
TN + FP

123
112 W. Ayadi et al.

Table 1 Summary of data

exploited in our study Tumor category Number of patients Number of slices View Number of slices

Meningioma 82 708 Transverse 209

Sagittal 231
Coronal 268
Pituitary 62 930 Transverse 291
Sagittal 320
Coronal 319
Glioma 89 1426 Transverse 494
Sagittal 495
Coronal 437

Fig. 2 Some examples of brain tumors: a, b present a glioma in several subjects. c presents a meningioma, and d presents a pituitary tumor from
different subjects [61]

vectors are concatenated in one row, which represents the

characteristic vector for this image. The first vector length
is about 9600, while the second is about 8100 features. This
vector is used as input for Support Vector Machine with a
linear kernel.
Table 2 outlines the obtained results for a different
approach using the accuracy of fivefold CV. Based on results,
the concatenation of two techniques of feature extraction can
lead to the highest accuracy, which reaches 90.27%. The
highest results are detailed below.
With the aim to enhance the MR images, we have uti-
lized the normalization step. The proposed system exploited
DSURF and HoG to extract features from each image, which
will be concatenated and become an input vector for SVM
Fig. 3 Fivefold CV scheme with a linear kernel. The confusion matrix is calculated for
all tumors category, and it is detailed in Table 3.
(CV) in aim to gain effective classification performance. The Table 3 shows the confusion matrix. From this table,
233 patients were chosen randomly into five subsets to check we can note that the proposed technique classifies 2766
the robustness of the suggested technique. The algorithm cases correctly and 298 cases incorrectly. It yields accept-
used four files for training, while it is tested with the residual able results, which attain 90.27% as the whole accuracy. We
file as explained in Fig. 3. can note also that the meningioma tumor class gains the high-
For each image, the normalization process is applied. The est misclassifications rate. The misclassification is due to the
feature extracted by DSURF is reshaped into a vector with fewer images number furnished for this class in the original
1 row. Also for the feature extracted by HoG. Then the two dataset.

123
Brain tumor classification based on hybrid approach 113

Table 2 Brain tumor classification accuracy using different approaches of the model. Table 4 summarizes the classification perfor-
Preprocessing Feature extraction Classifier Accuracy mance of each tumor type.
From Table 4, we note that the proposed model provides
− HoG SVM LIN 86.42 90.83%, 92.66%, and 97.06% as accuracy for Meningioma,
Median filter HoG SVM LIN 86.65 Glioma and Pituitary tumor, respectively. The sensitivity is
Normalization HoG SVM LIN 86.04 another factor of classification performance, which attains,
− DSURF SVM LIN 89.58 respectively, 84.89%, 88.85%, and 96.56%. For all the
Median filter DSURF SVM LIN 89.26 classes, the specificity values are high which means that the
Normalization DSURF SVM LIN 89.58 model correctly recognizing the samples without a specific
− HoG + DSURF SVM LIN 90.11 disease.
Median filter HoG + DSURF SVM LIN 89.78 To further test the new scheme effectiveness, a comparison
Normalization HoG + DSURF SVM LIN 90.27 with several previous methods using the same data has been
(−) indicates that no preprocessing is used made using average accuracy. Table 5 outlines a comparison
of results through the accuracy for various techniques. In
Table 3 Confusion matrix for the suggested technique the new schemes, the deep learning was exploited to classify
three types of tumor.
True/Auto Meningioma Glioma Pituitary
Each one suggested a new architecture to get an acceptable
Meningioma 601 54 53 accuracy result. Here, our suggested scheme provides the
Glioma 154 1267 5 highest accuracy, which reaches 90.27%. The gained results
Pituitary 20 12 898 use a strict statistical analysis (fivefold CV with 10 repe-
titions), which proves the reliability and robustness of the
proposed scheme.
Moreover, the computational time presents another factor
to a classification method evaluation. Here, the time of the test
stage is only considered. A brain MR image of size 512 * 512
needs about 0.27 s as overall computation time, which is
sufficiently fast.
Obtained results using a huge brain dataset proved the
success of the suggested approach. We can believe that the
suggested technique can be generalized also to various MRI
classification. It is proved that the concatenation of two fea-
ture extraction schemes increases the performance.
Several previous works [32] used the deep learning tech-
nique. It provides good accuracy, which attains 98%. The
Fig. 4 Comparison of several classification methods (%) long processing time represents the main disadvantage of this
technique that needs about 2 h 40 min. The proposed method
has a lower time complexity compared to other systems. It
Moreover, to prove the efficacy of the proposed scheme, needs about 0.27 s as overall computation time to classify an
accuracy comparison is made against various classifiers like image with a size of 512 * 512, which is sufficiently fast.
KNN with various values of K, Random Tree (RT), and SVM. Other systems proposed in the literature [28] exploit a
The extracted features are used as the input of these classi- region segmentation for MRI image before using it to the
fiers. Fivefold cross-validation was exploited in the aim to proposed method. The suggested technique uses the whole
validate the classification results. Figure 4 shows that with image and without any segmentation technique.
various values of K, the KNN classifier can reach about 83% Other authors [33–35] select a few images from the orig-
as accuracy. The Rtrees classifier is slightly better efficiency inal dataset with the aim to gain a higher classification
than the KNN classifier and can achieve 84.72% as accu- accuracy.
racy. SVM classifier gained the best accuracy, which attains Due to the various operations in the CNN layers, the
89.88% for SVM with POLY kernel and 90.27% for SVM DL requires high computational power (GPUs) compared to
with LIN kernel. traditional machine learning which requires a lower com-
Generally, many previous works utilized accuracy to eval- putational cost. In the presence of limited computational
uate model performance. In this study, we have exploited resources (CPU), we exploit the machine learning for brain
other measures like sensitivity, specificity, precision, and F1- tumor classification in this work.
score, in order to have a clear indicator of the generalization

123
114 W. Ayadi et al.

Table 4 Performance of the

proposed method on the test Class Accuracy Sensitivity Specificity Precision F1-score
dataset
Meningioma 90.83 84.89 92.61 77.55 81.05
Glioma 92.66 88.85 95.97 95.05 91.84
Pituitary tumor 97.06 96.56 97.28 93.93 95.23

Table 5 Performance
comparison with several Method Number of images Best accuracy Manual Evaluation method
classification approaches used segmentation

Abiwinanda et al. 2100 (700 from 84.19 No Train-validation

[35] Different each tumor type)
ConvNets
Paul et al. [34] 989 (axial only) 84.52 No Fivefold
ConvNet, 64 ×
64
Tahir et al. [36] 3064 86 No Tenfold
SVM +
Preprocessing
Afshar et al. [26] 3064 86.56 using Both Not mentioned
CapsNet segmentation
72.13 using raw
images
Paul et al. [34] 989 (axial only) 90.26 No Fivefold
ConvNet, 256 ×
256
Proposed 3064 90.27 No Fivefold
Cheng et al. [28] 3064 87.54 Yes Fivefold
Intensity
histogram
GLCM 89.72
BoW 91.28
Deepak et al. [32] 3064 92.3 No Fivefold
CNN
CNN + SVM 97.8
CNN + KNN 98
Ahmad et al. [33] 510 99.3 No Train-validation
DWT + SVM

7 Discussion In this work, we use various measures in the aim to gain a

clear indicator of the generalization of the proposed model.
In this paper, we have suggested a new technique for MRI As shown in Table 4, the suggested technique provides good
brain tumor classification. It exploits DSURF and HoG with accuracy, which reaches 90.83% for Meningioma, 92.66%
the aim to extract good features. SVM is used as a classifier for Glioma, and 97.06% for Pituitary tumor. The overall accu-
in the last step. The experimental study of the proposed tech- racy is about 90.27%. Furthermore, the obtained results are
nique is carried out based on the publicly Figshare dataset. compared with some new previous works. As detailed in
It contains 3064 images correspond to meningioma, glioma, Table 5, it is clear that our model achieved good accuracy
and pituitary tumors. The classification results are analyzed compared to some previous works.
and various evaluation metrics are calculated as accuracy, Based on the obtained results, the proposed model gained
sensitivity, specificity, precision, and F1-score. good accuracy in MRI brain tumor classification using
In general, only the accuracy is exploited in the major- this dataset. However, the MRI brain images classification
ity of the previous works to evaluate the proposed scheme. represents a challenging problem owing to the variety in
However, the use of the accuracy only for comparison can intensities, size, orientation, and shape. In addition to image
be misleading, since it ignores the sensitivity to imbalanced contrast and the noise perturbations. Furthermore, the med-
data. In this case, some class performances can become better ical datasets exploited are often limited in size and hard to
than others. access.

123
Brain tumor classification based on hybrid approach 115

In our opinion, the suggested scheme is effective to clas- 2. de Robles, P., et al.: The worldwide incidence and prevalence of pri-
sify MRI brain tumors as type, which help doctors to take mary brain tumors: a systematic review and meta-analysis. Neuro
Oncol. 17(6), 776–783 (2015)
the precise decision in a short time. It is simple and easy to 3. Tamimi, A.F., et al.: Epidemiology of malignant and non-
apply with a low cost. We believe also that our model can malignant primary brain tumors in Jordan. Neuroepidemiology
be exploited to classify other tumor types as breast cancer, 45(2), 100–108 (2015)
lung cancer, liver cancer,etc. and other medical image types 4. Saba, T., et al.: Annotated comparisons of proposed preprocess-
ing techniques for script recognition. Neural Comput. Appl. 25(6),
as ultrasound, X-ray, etc. 1337–1347 (2014)
5. Segal, R.L., Miller, K.D., Jemal, A.: Cancer statistics, 2018. CA
Cancer J. Clin. 68(1), 7–30 (2018)
6. Papageorgiou, T.S., et al.: Diffusion tensor imaging in brain tumors:
8 Conclusion a study on gliomas and metastases. Physica Med. 31(7), 767–773
(2015)
A new brain tumor classification technique is suggested. 7. Tsougos, I., et al.: The importance of MR multiparametric differen-
tial diagnosis. FASMA. A clinical decision support system. Physica
The suggested technique was applied to classify a cerebral
Medica Eur. J. Med. Phys. 30, 38 (2014)
tumor into three categories (glioma, meningioma, and pitu- 8. Chen, J.-W., Zhou, C.-F., Lin, Z.-X.: The influence of different
itary tumor). Various experiments are realized based on a classification standards of age groups on prognosis in high-grade
benchmark dataset to test the system. The obtained results hemispheric glioma patients. J. Neurol. Sci. 356(1–2), 148–152
(2015)
illustrated that our method provides perfect performance than
9. Nabizadeh, N., Kubat, M.: Brain tumors detection and segmenta-
other recent schemes. The main work of this study lies in tion in MR images: Gabor wavelet vs. statistical features. Comput.
three steps. Firstly, we apply normalization to ameliorate Electr. Eng. 45, 286–301 (2015)
MR images. Secondly, we extract DSURF and HoG fea- 10. Shanthakumar, P., Ganeshkumar, P.: Performance analysis of clas-
sifier for brain tumor detection and diagnosis. Comput. Electr. Eng.
ture, which are sufficient to accomplish better performance.
45, 302–311 (2015)
Finally, SVM with a linear kernel is exploited as a classi- 11. Zhang, Y., et al.: Detection of Alzheimer’s disease and mild cogni-
fier. The suggested system achieves 90.27% as accuracy. The tive impairment based on structural volumetric MR images using
radiologist can use this algorithm as a second viewpoint. 3D-DWT and WTA-KSVM trained by PSOTVAC. Biomed. Signal
Process. Control 21, 58–73 (2015)
It helps to minimize computational time and providing an
12. Kalbkhani, H., Shayesteh, M.G., Zali-Vargahan, B.: Robust algo-
improved accuracy rate. The automatic classification system rithm for brain magnetic resonance image (MRI) classification
will reduce enormously the manpower and the time needed based on GARCH variances series. Biomed. Signal Process. Con-
for brain tumor diagnosis. Thus, it is conducted to minimize trol 8(6), 909–919 (2013)
13. Wang, S., et al.: Feed-forward neural network optimized by
the false classification rate.
hybridization of PSO and ABC for abnormal brain detection. Int.
As future work, we will exploit images from several J. Imaging Syst. Technol. 25(2), 153–164 (2015)
modalities to add robustness to the proposed scheme. The 14. Sachdeva, J., et al.: A package-SFERCB-“segmentation, feature
suggested method can be validated also through modern extraction, reduction and classification analysis by both SVM and
ANN for brain tumors”. Appl. Soft Comput. 47, 151–167 (2016)
machine learning approaches as dictionary learning and deep
15. Das, S., Chowdhury, M., Kundu, M.K.: Brain MR image clas-
learning. sification using multiscale geometric analysis of ripplet. Prog.
Electromagn. Res. 137, 1–17 (2013)
16. Gordillo, N., Montseny, E., Sobrevilla, P.: State of the art survey
Compliance with ethical standards on MRI brain tumor segmentation. Magn. Reson. Imaging 31(8),
1426–1438 (2013)
17. Jiang, J., et al.: 3D brain tumor segmentation in multimodal MR
Conflict of interest The authors certify that they have no affiliations with images based on learning population-and patient-specific feature
or involvement in any organization or entity with any financial inter- sets. Comput. Med. Imaging Graph. 37(7–8), 512–521 (2013)
est (such as honoraria; educational grants; participation in speakers’ 18. Wu, Y., et al.: Semi-automatic segmentation of brain tumors using
bureaus; membership, employment, consultancies, stock ownership, or population and individual information. J. Digit. Imaging 26(4),
other equity interest; and expert testimony or patent-licensing arrange- 786–796 (2003)
ments) or non-financial interest (such as personal or professional 19. Dogra, J., Jain, S., Sood, M.: Glioma extraction from MR images
relationships, affiliations, knowledge or beliefs) in the subject matter employing gradient based kernel selection graph cut technique. Vis.
or materials discussed in this manuscript. All of the authors have read Comput. 36(5), 875–891 (2020)
and approved the paper, and it has not been published previously nor is 20. Rahim, M.S.M., et al.: 3D bones segmentation based on CT images
it being considered by any other peer-reviewed journal. visualization. Biomed. Res. (0970-938X) 28(8), 3641–3644 (2017)
21. Rahim, M.S.M., et al.: Ear biometrics for human classification
based on region features mining. Biomed. Res. (0970-938X)
28(10), 4660–4664 (2017)
References 22. Nagpal, J., Vidyarthi, A., Mittal, N.: CLOM: counting label
occurrence matrix for feature extraction in MR images. In: 2015
1. Iqbal, S., et al.: Computer-assisted brain tumor type discrimination International Conference on Signal Processing and Communica-
using magnetic resonance imaging features. Biomed. Eng. Lett. tion (ICSC), pp. 216–221. IEEE (2015)
8(1), 5–28 (2018)

123
116 W. Ayadi et al.

23. Al-Shaikhli, S.D.S., Yang, M.Y., Rosenhahn, B.: Brain tumor 44. Komorkiewicz, M., Kluczewski, M., Gorgon, M.: Floating point
classification using sparse coding and dictionary learning. In: HOG implementation for real-time multiple object detection. In:
2014 IEEE International Conference on Image Processing (ICIP), 22nd International Conference on Field Programmable Logic and
pp. 2774–2778. IEEE (2014) Applications (FPL), pp. 711–714. IEEE (2012)
24. Hang, X., Wu, F.-X.: Sparse representation for classification of 45. Feng, R., Du, Q., Li, X., et al.: Robust registration for remote sens-
tumors using gene expression data. J. Biomed. Biotechnol. 2009, ing images by combining and localizing feature-and area-based
1–6 (2009) methods. ISPRS J. Photogram. Remote Sens. 151, 15–26 (2019)
25. Li, X., Shen, H., Zhang, L., et al.: Recovering quantitative remote 46. Xi, P., Guan, H., Shu, C., et al.: An integrated approach for med-
sensing products contaminated by thick clouds and shadows using ical abnormality detection using deep patch convolutional neural
multitemporal dictionary learning. IEEE Trans. Geosci. Remote networks. Vis. Comput. 35, 1–14 (2019)
Sens. 52(11), 7086–7098 (2014) 47. Ayadi, W., Elhamzi, W., Charfi, I., et al.: A hybrid feature extrac-
26. Afshar, P., Mohammadi, A., Plataniotis, K.N.: Brain tumor type tion approach for brain MRI classification based on Bag-of-words.
classification via capsule networks. In: 2018 25th IEEE Interna- Biomed. Signal Process. Control 48, 144–152 (2019)
tional Conference on Image Processing (ICIP), pp. 3129–3133. 48. Ergin, S., Kilinc, O.: A new feature extraction framework based
IEEE (2018) on wavelets for breast cancer diagnosis. Comput. Biol. Med. 51,
27. Liu, Y.-h., et al.: Classification of MR tumor images based on Gabor 171–182 (2014)
wavelet analysis. J. Med. Biol. Eng. 32(1), 22–28 (2012) 49. Zhi, R., Liu, M., Zhang, D.: A comprehensive survey on automatic
28. Cheng, J., et al.: Enhanced performance of brain tumor classifi- facial action unit analysis. Vis. Comput. 36(5), 1067–1093 (2020)
cation via tumor region augmentation and partition. PLoS ONE 50. Srisamosorn, V., Kuwahara, N., Yamashita, A., et al.: Human
10(10), e0140381 (2015) position and head direction tracking in fisheye camera using ran-
29. Hemanth, D.J., et al.: Performance improved iteration-free artifi- domized ferns and fisheye histograms of oriented gradients. Vis.
cial neural networks for abnormal magnetic resonance brain image Comput. 36, 1–14 (2019)
classification. Neurocomputing 130, 98–107 (2014) 51. Fazl-Ersi, E., Nooghabi, M.K.: Revisiting correlation-based fil-
30. Caulo, M., et al.: Data-driven grading of brain gliomas: a multi- ters for low-resolution and long-term visual tracking. Vis. Comput.
parametric MR imaging study. Radiology 272(2), 494–503 (2014) 35(10), 1447–1459 (2019)
31. Lin, B.-J., et al.: Correlation between magnetic resonance imaging 52. Cortes, C., Vapnik, V.: Support-vector networks. Mach. Learn.
grading and pathological grading in meningioma. J. Neurosurg. 20(3), 273–297 (1995)
121(5), 1201–1208 (2014) 53. Kaashki, N.N., Safabakhsh, R.: RGB-D face recognition under var-
32. Deepak, S., et al.: Brain tumor classification using deep CNN fea- ious conditions via 3D constrained local model. J. Vis. Commun.
tures via transfer learning. Comput. Biol. Med. 111, 103345 (2019) Image Represent. 52, 66–85 (2018)
33. Sarhan, A.M.: Brain tumor classification in magnetic resonance 54. Tellez, E.S., et al.: An automated text categorization framework
images using deep learning and wavelet transform. J. Biomed. Sci. based on hyperparameter optimization. Knowl. Based Syst. 149,
Eng. 13(6), 102 (2020) 110–123 (2018)
34. Paul, J.S., et al.: Deep learning for brain tumor classification. In: 55. Wang, H., et al.: A support vector machine-based ensemble algo-
Medical Imaging 2017: Biomedical Applications in Molecular, rithm for breast cancer diagnosis. Eur. J. Oper. Res. 267(2),
Structural, and Functional Imaging, vol. 10137, p. 1013710 (2017) 687–699 (2018)
35. Abiwinanda, N., et al.: Brain tumor classification using convolu- 56. Shanker, R., Bhattacharya, M.: An automated computer-aided
tional neural network. In: World Congress on Medical Physics and diagnosis system for classification of MR images using texture fea-
Biomedical Engineering 2018, pp. 183–189. Springer, Singapore tures and gbest-guided gravitational search algorithm. Biocybern.
(2019) Biomed. Eng. 40(2), 815–835 (2020)
36. Tahir, B., et al.: Feature enhancement framework for brain 57. Toğaçar, M., Ergen, B., Cömert, Z.: Detection of lung cancer on
tumor segmentation and classification. Microsc. Res. Tech. 82(6), chest CT images using minimum redundancy maximum relevance
803–811 (2019) feature selection method with convolutional neural networks. Bio-
37. Li, X., Feng, R., Guan, X., et al.: Remote sensing image mosaick- cybern. Biomed. Eng. 40(1), 23–39 (2020)
ing: achievements and challenges. IEEE Geosci. Remote Sens. 58. Khosla, A., Khandnor, P., Chand, T.: A comparative analysis of sig-
Mag. 7(4), 8–22 (2019) nal processing and classification methods for different applications
38. Wang, J.-G., et al.: Boosting dense SIFT descriptors and shape based on EEG signals. Biocybern. Biomed. Eng. 40(2), 649–690
contexts of face images for gender recognition. In: 2010 IEEE (2020)
Computer Society Conference on Computer Vision and Pattern 59. Ain, Q., Jaffar, M.A., Choi, T.-S.: Fuzzy anisotropic diffusion based
Recognition-Workshops, pp. 96–102. IEEE (2010) segmentation and texture based ensemble classification of brain
39. Bay, H., et al.: Speeded-up robust features (SURF). Comput. Vis. tumor. Appl. Soft Comput. 21, 330–340 (2014)
Image Underst. 110(3), 346–359 (2008) 60. Gupta, T., et al.: Classification of patients with tumor using MR
40. Kashif, M., et al.: Feature description with SIFT, SURF, BRIEF, FLAIR images. Pattern Recogn. Lett. 139, 1–6 (2017)
BRISK, or FREAK? A general question answered for bone age 61. Huang, M., et al.: Content-based image retrieval using spatial lay-
assessment. Comput. Biol. Med. 68, 67–75 (2016) out information in brain tumor T1-weighted contrast-enhanced MR
41. Korkmaz, S.A., Binol, H.: Classification of molecular structure images. PLoS ONE 9(7), 102754 (2014)
images by using ANN, RF, LBP, HOG, and size reduction methods
for early stomach cancer detection. J. Mol. Struct. 1156, 255–263
(2018)
Publisher’s Note Springer Nature remains neutral with regard to juris-
42. Shim, J.-S., Ju, Y.-W., Park, S.-C.: Design and implementation of a
dictional claims in published maps and institutional affiliations.
Pedestrian recognition algorithm using trilinear interpolation based
on HOG-UDP. J. Supercomput. 74(2), 787–800 (2018)
43. Liu, B., Wu, H., Su, W., et al.: Rotation-invariant object detection
using Sector-ring HOG and boosted random ferns. Vis. Comput.
34(5), 707–719 (2018)

123
Brain tumor classification based on hybrid approach 117

Wadhah Ayadi is a PhD student Wajdi Elhamzi is a member of

at the University of Monastir the Laboratory of Electronics
(Tunisia) and a member of the & Microelectronics at Monastir
Laboratory of Electronics & University (Tunisia). He received
Microelectronics, and his topic a PhD degree in electronic and
focused on medical imaging microelectronic from University
based on machine and deep learn- of Dijon (France) and University
ing and hardware design-based of Monastir (Tunisia) in 2013.
FPGA. He has been involved in FPGA-
based hardware implementation
of image processing algorithms,
medical imaging, deep learning,
and cryptography.

Imen Charfi received his PhD Mohamed Atri is Professor in the

degree in electronic and micro- Department of Computer Engi-
electronic from University of neering at College of Computer
Dijon (France) and University Science, King Khalid University,
of Monastir (Tunisia) in 2013. Saudi Arabia. He received his
His topic focused on medical on Ph.D. Degree in Microelectronics
machine and deep learning and from the University of Monastir,
hardware design-based FPGA. Tunisia in 2001. He has obtained
the HDR degree from the Uni-
versity of Monastir, Tunisia in
2011. He is currently a member
of the Laboratory of Electronics
& Microelectronics, Faculty of
Science of Monastir. His research
includes Circuit and System
Design, Image processing, Network Communication, IPs and SoC.

123