Abstract: Medical image analysis is a very interesting research area, and it is a significant challenge for
researchers. Due to the complexity of the brain structure, accurate diagnosis of brain tumors is extremely difficult.
In recent years, research focused on medical image processing to solve this problem by relying on deep learning
techniques, and it has achieved good results in this field. This paper proposes an efficient convolutional neural
network model for MR brain image segmentation and analysis. The novel model consists of segmentation
efficient-CNN and pre-efficient-CNN blocks for dataset diminution and improvement blocks. The unique
efficient-CNN is specially designed according to the model proposed by ASCNN (application) CNN-specific) to
perform unidirectional and transverse feature extraction and tumor and pixel classification. The recommended
Full-ReLU activation feature halves the number of cores in a high-coil filtered winding layer without reducing
process quality. In this specific efficient-CNN consists of 8 convolutional layers and 110 kernels. The experiment
results were done using the MR brain database from the Arizona university, including eluding with and without
tumor images. The proposal model achieved an accuracy of 97.2% to 98%, which proves the efficiency of the
model and its ability to assist in the early diagnosis of brain tumors with sufficient accuracy to support the doctors'
decision during diagnosis.
Key Words: - Brain Tumor, MRI databases, Medical Image, CNN Model and accuracy.
Received: May 11, 2021. Revised: March 5, 2022. Accepted: April 7, 2022. Published: May 5, 2022.
1 Introduction
The brain is one of the most complex mechanisms in
humans. The body is composed of billions of cells.
Brain tumors can be expressed as follows:
Tissues that are in places that should not be in our
brains, or each substance grows uncontrollably where
it should not be. Early and correct detection of brain
tumors is essential in this cancer species; it is a killer
[1]. In this work, the classification process was
performed using MRI images. The network must
form an extensive network before performing this
procedure. In this article, deep learning methods can
yield positive results in large databases. Thanks to
computer-aided systems, experts can diagnose the
disease. In this way, errors with traditional methods
can be avoided [2]. There are studies in the literature
that use different models and architectures [3].
Isselmou et al. [4] proposed a hybrid convolutional
neural network with fuzzy c-means algorithm for
brain tumor detection and analysis. Their proposed
model combined a new deep learning algorithm with
the traditional method. Their model provides good
results in brain tumor detection using MRI images
and good performance during the analysis step.
Isselmou et al. [5] suggested a differential
convolutional neural network model for brain tumor
classification using MR big data. The new
differential model innovative is added two layers and
An Efficient Convolutional Neural Network Model for Brain
MRI Segmentation
1ISSELMOU ABD EL KADER, GUIZHI XU, 1ZHANG SHUAI, 2EL MAALOUMA SIDI
BRAHIM, 1SANI SAMINU
1School of Health Science and Biomedical Engineering, Hebei University of Technology.
Tianjin City, CHINA.
2College of Engineering, Zhejiang Normal University, Jinhua City, CHINA.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.10
Isselmou Abd El Kader, Guizhi Xu,
Zhang Shuai, El Maalouma Sidi Brahim, Sani Saminu
E-ISSN: 2224-2902
77
Volume 19, 2022
one operator in the original architecture of the
convolutional neural network. The model has shown
excellent classification of the brain tumors as yes/no
tumor using massive data and obtained outstanding
overall achievement during training and testing
stages without any technical problems or data
balance.
Guizhi Xu et al. [6] created a deep wavelet transform
model for brain tumor detection and classification
using different MRI BRATS databases. The model
uses 4-connected for thresholding cluster pixels in
the input MR databases, two layers to provide slices
images segmentation and 200 units in the first layer
and 400 units in the second layer. The model gives
excellent results and ability about brain classification
and segmentation data and good performance
analysis using FNR and FPR values.
Guizhi Xu et al. [7] suggested a hybrid convolutional
neural network with a deep watershed auto-encoder
for brain tumor detection and analysis. The hybrid
model training and testing using BRATS MR
extensive databases. The model is based on a
complex matrix combined convolutional neural
network and a deep watershed auto-encoder to
provide excellent detection and classification of big
data achieved excellent performance using different
values.
Ahmet Çinar et al. [8] proposed a hybrid
convolutional neural network architecture for
brain tumor detection on MR images. The model
used resnet50 architecture. They removed the
last layers of the Resnet50 model and added 8
new layers. The model produced an excellent
performance based on accuracy value.
Khan, M.A ET at [9] presents multiple
automated models based on deep learning for
brain tumor classification using T1-T2 weight
and FLAIR images. The model includes five
steps; In the first step; is applying linear contrast
stretching based on edge-based histogram
equalization and discrete cosine transform
(DCT). In the second step, they performed deep
learning feature extraction by using transfer
learning, two pre-trained convolutional neural
networks (CNN) models, namely VGG16 and
VGG19. The third step implemented a
correntropy-based joint learning approach with
the extreme learning machine (ELM) to choose
the best features. In the fourth step, they fused
the partial least square (PLS)-based robust
covariant features in one matrix. The fed to ELM
for the last classific in the fifth station. In the fifth
step, The multiple models obtained good
performance based on accuracy value.
Tanzila Saba et al. [10] proposed a grab cut
model used to precisely distribute actual lesion
symptoms. At the same time, the VGG-19
optical engineering set is tuned explicitly for
function and then sequenced by a sequence-
based approach to manual features (shape and
texture). These characteristics are optimized by
entropy for accurate and fast classification and
provide fusion vectors to the classifier. The
model was tested using medical image
computing and computer-assisted intervention
(MICCAI) database and obtained good
performance based on dice similarity coefficient
(DSC) value achievement.
Hemanth, D.J et al. [11] presented a modified
deep convolutional neural network model for
brain tumor classification. The second objective
of the work is to reduce the complexity of the
original convolutional neural network
architecture. Suitable, the training algorithm is
adjusted to reduce the number of parameter
adjustments. The proposed modifications
method eliminates changing the weight in the
fully connected layer. Instead, use a simple
mapping process to find the weights of this fully
connected layer. Thus, the proposed method
considerably reduces the complexity of the
calculation. The model training uses different
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.10
Isselmou Abd El Kader, Guizhi Xu,
Zhang Shuai, El Maalouma Sidi Brahim, Sani Saminu
E-ISSN: 2224-2902
78
Volume 19, 2022
MR brain abnormal images and gives a good
performance.
2 Databases or Materials
This paper used 6000 MR brain images from the
Arizona university website for the brain database.
The database consists of 4000 MR brain images with
tumor, and 2000 MR brain images without tumor
figure 1 represent sample with tumor, and figure 2
denotes sample without tumor.
Figure 1: Sample of MR Brain Images with Tumor.
Figure 2: Sample of MR Brain Images with Tumor.
3 Method of the Proposed Model
Each MR image is displayed as input for a hash
system with four different 3D MRI images road.
Because all 3D image voxels are cut into 2D
segments, which become images of 3D pixels in a 2D
chip. In the system proposed by efficient CNN, brain
tumors with 3D images are fragmented. This is done
by dividing these 2D slides.
The proposed model uses the most important data
partition that occurs in CNNs. There are efficient
CNN and post-efficient CNN blocks to reduce CNN,
designed to improve CNN computing efficiency. In
addition, by doing so, the system will have a higher
degree of certainty, less Randomness, and better
replication. The final step is output consisting of
segmentation, classification slices results, and the
overall performance of the efficient-CNN model—
the stages of the proposed efficient-CNN model is
illustrated in figure 3.
Figure 3: Schematic explaining the stages of the proposed
efficient-CNN model.
3.1 Pre-efficient-CNN
The 3D images entered from commonly used
datasets, such as the University of Arizona Brain Site
Database, are 240×240×155 in size and are generated
through late-× acquisition registrations. Predictive
CNN's simplify CNN calculations by reducing the
amount of 3D image data. Specifically, this is done
by detecting and removing tumor-free fragments
from brain images.
(1)
Where
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.10
Isselmou Abd El Kader, Guizhi Xu,
Zhang Shuai, El Maalouma Sidi Brahim, Sani Saminu
E-ISSN: 2224-2902
79
Volume 19, 2022
3.2 Convolutional Neural Network (CNN)
Given that the results proposed by CNN are an
essential function of retail, the quality of the
treatment determines the overall performance of the
whole system. In order not to compromise the quality
of processing while reducing the amount of
computation, CNNs must be designed on demand,
not just to support and modify existing networks. All
calculation elements are necessary and fair enough to
process input data for specific brain fragmentation
tasks. In addition, a new activation feature called
Full-ReLU (Fully Corrected Linear Unit) is used in
some layers to help create it.
More efficient computing. The new Full-ReLU
activation feature is presented in section 3.2.1.
3.2.1 Full-ReLU – a novel activation function
The activation feature performs a nonlinear
transformation in the wrapper layer. ReLU
(Correction Linear Unit), defined as
is probably the most commonly used
activation function in CNN designs. However, they
are not perfect. In the case of high eddy current
filtration, a single escape process can produce a series
of positive and negative elements that represent
signal differences in opposite directions. If ReLU
applies to these projects, negatives will be
eliminated, resulting in a loss of information.
It is derived from a data set in two groups 2,
namely and and it is mathematically
expressed as follows:
(2)
3.3 Post Efficient-CNN
After being ranked by CNN, the post-CNN block is
placed Identifies pixels that have been incorrectly
classified as positive. Identification depends on
whether the brain tumor and its tumor amplify basic
(if any) is a 3D object, and the surface of each entity
must be Found in several consecutive slides.
The thickness of the entire tumor that can be detected
is considered at least 1/25 of the diameter of the brain.
If the 3D brain image consists of 150 slices, this
thickness corresponds to at least seven consecutive
cuts. As a whole, tumor areas appear in less than
seven successive segments, with pixels in this area
can be misclassified and reclassified as tumor-free
pixels.
In short, efficient CNNs are explicitly based on
Features brain images, along with pre-and after
CNN's functions accurate and effective blocking and
fragmentation of brain tumors he has a very low
account cost.
4 Experimental Results and Analysis
In this work, the experiments were done on a Jupiter
notebook environment using Lenovo workstation
16G. The experiments results consist following steps:
4.1 MR Databases processed
.
Figure 4: Results of MR Database processing based on
efficient-CNN Model.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.10
Isselmou Abd El Kader, Guizhi Xu,
Zhang Shuai, El Maalouma Sidi Brahim, Sani Saminu
E-ISSN: 2224-2902
80
Volume 19, 2022
In the MR database process section, we presented the
result of brain tumor detection using the efficient-
CNN model in figure 4. It shows the ability of the
proposed model to segment MR images and detect
the area of tumor in each image.
Figure 5: Results of classification as yes Tumor based on
Efficient-CNN Model.
Figure 5 denotes the efficient-CNN model's
classification results as yes tumor commend. The
process gives excellent results and proves the
proposed model's capacity to classify the MR
database as a brain tumor dataset.
Figure 6: Results of classification as No Tumor based on
Efficient-CNN Model.
Figure 6 represents the results of classification based
on the efficient-CNN model as no tumor commend.
The results have shown good classification results as
MR database without tumor. From the section on MR
databases, we can see the great abilities of the
proposed model during segmentation, detection
Tumor, and classification of no or yes tumor.
4.2 Evaluation of Efficient-CNN Model
The evaluation of the efficient-CNN model is
based on a comparison with five existing machine
learning models such as CNN, DNN, ANN, KNN,
and Multi-SVM, using the following values:
accuracy, sensitivity, dice score, and false discovery
rate (FDR) values.
(3)
(4)
(5)
(6)
Figure 7 denotes the accuracy value results of the
proposed model, and it shows a good achievement
during the training and testing steps.
Figure 7: Results of accuracy value using Efficient-CNN
Model
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.10
Isselmou Abd El Kader, Guizhi Xu,
Zhang Shuai, El Maalouma Sidi Brahim, Sani Saminu
E-ISSN: 2224-2902
81
Volume 19, 2022
Figure 8 represents the loss validation value results
of a proposed model during training and testing steps,
and it achieved an excellent loss value result.
Figure 8: Results of loss validation value using Efficient-
CNN Model.
Table .1 Comparison between Efficient-CNN Model
Performance with existing Model.
Model
Accuracy%
Sensitivity%
Dice
Score
FDR
CNN
[12]
94
92.5
0.46
0.52
DNN
[13]
91.4
93.2
0.51
0.61
ANN
[14]
88
62
0.68
0.56
KNN
[15]
78
46
0.55
0.52
Multi-
SVM
[16]
84
55
0.6
0.56
Efficient-
CNN
97.2
95.5
0.31
0.47
Table 1 compares the efficient-CNN model with
existing models such as CNN, DNN, ANN, KNN,
and Multi-SVM using four metrics values. The
convolutional neural network (CNN) model
produced an accuracy of 94%, sensitivity of 92.5%,
dice score of 0.46, and FDR of 0.52. The deep neural
network model obtains an accuracy of 91.4%,
sensitivity of 93.2%, dice score of 0.51, and FDR of
0.61. The artificial neural network model achieved an
accuracy of 88%, sensitivity of 62%, dice score of
0.68, and FDR of 0.56. K-neural network model
gives an accuracy of 78%, sensitivity of 46%, dice
score of 0.55, and FDR of 0.52. The multi-SVM
model obtains an accuracy of 84%, a sensitivity of
55%, a dice score of 0.6, and an FDR OF 0.56. The
efficient-CNN model gives an accuracy of 97.7.2%,
a sensitivity of 95.5%, a dice score of 0.31, and an
FDR of 0.47. According to an analysis of existing
model performance using accuracy, sensitivity, dice
score, and false discovery rate (FDR), the efficient-
CNN model achieved better overall performance than
CNN, DNN, ANN, KNN, and Multi-SVM models.
5 Conclusion
Due to the excellent achievement of the deep learning
model in the medical image analysis field, we
proposed an efficient-CNN model with a new
structure and innovated for brain tumor detection and
classification with high performance. The proposed
model shows excellent results during the MR
database process, detects the tumor area in different
images, and classifies the database as yes/no tumor
effectiveness (see section 4.1). The evaluation stage
of the proposed model shows excellent overall
performance, such as accuracy of 97.2 to 98%, a
sensitivity of 95.5%, and a dice score of 0.31 and
FDR of 0.47, which are better than existing models.
Based on the high performance of the efficient-CNN
model, we strongly recommend using it as a
computer brain aide technique for early brain tumor
detection and reducing the number of deaths. We
will add more layers in future work and make the
model deeper.
References:
[1] Prabhu LAJ, Jayachandran A. Mixture model
segmentation system for parasagittal
meningioma brain tumor classification based on
hybrid feature vector. J Med Syst 2018; 42 (12):
pp 251.
[2] Arasi PRE, Suganthi M. A clinical support
system for brain tumor classification using soft
computing techniques. J Med Syst 2019; 43(5):
pp 144.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.10
Isselmou Abd El Kader, Guizhi Xu,
Zhang Shuai, El Maalouma Sidi Brahim, Sani Saminu
E-ISSN: 2224-2902
82
Volume 19, 2022
[3] Yildirim M, Çinar A. Classification of white
blood cells by deep learning methods for
diagnosing disease. Revue d'Intelligence
Artificial 2019; 33(5): pp 335–40.
[4] G. X. Abd El Kader Isselmou, Z. Shuai, S.
Saminu, I. Javaid, and I. S. Ahmad, "Brain
Tumor identification by Convolution Neural
Network with Fuzzy C-mean Model Using MR
Brain Images." INTERNATIONAL JOURNAL
OF CIRCUITS, SYSTEMS AND SIGNAL
PROCESSING, 2020, 14(14), pp 1096-1102.
[5] I. Abd El Kader, G. Xu, Z. Shuai, S. Saminu, I.
Javaid, and I. Salim Ahmad, "Differential deep
convolutional neural network model for brain
tumor classification," Brain Sciences,2021 vol.
11, pp. 352.
[6] I. Abd El Kader, G. Xu, Z. Shuai, S. Saminu, I.
Javaid, I. S. Ahmad, et al., "Brain Tumor
Detection and Classification on MR Images by
a Deep Wavelet Auto-Encoder Model,"
Diagnostics, 2021, vol. 11, pp. 1589.
[7] A. El Kader, G. Xu, Z. Shuai, and S. Saminu,
"Brain tumor detection and classification by
hybrid CNN-DWA model using MR images,"
Current Medical Imaging,2021 vol. 17, pp.
1248-1255.
[8] A. Çinar and M. Yildirim, "Detection of tumors
on brain MRI images using the hybrid
convolutional neural network architecture,"
Medical hypotheses, 2020, vol. 139, pp. 109-
684.
[9] M. A. Khan, I. Ashraf, M. Alhaisoni, R.
Damaševičius, R. Scherer, A. Rehman, et al.,
"Multimodal brain tumor classification using
deep learning and robust feature selection: A
machine learning application for radiologists,"
Diagnostics, 2020, vol. 10, pp. 565.
[10] T. Saba, A. S. Mohamed, M. El-Affendi, J.
Amin, and M. Sharif, "Brain tumor detection
using a fusion of handcrafted and deep learning
features," Cognitive Systems Research, 2020,
vol. 59, pp. 221-230.
[11] D. J. Hemanth, J. Anitha, A. Naaji, O. Geman,
and D. E. Popescu, "A modified deep
convolutional neural network for abnormal
brain image classification," IEEE Access, 2018
vol. 7, pp. 4275-4283.
[12] M. Sajjad, S. Khan, K. Muhammad, W. Wu, A.
Ullah, and S. W. Baik, "Multi-grade brain tumor
classification using deep CNN with extensive
data augmentation," Journal of computational
science,” 2019, vol. 30, pp. 174-182.
[13] S. K. Hasan and C. A. Linte, "A modified U-Net
convolutional network featuring a Nearest-
neighbour Re-sampling-based Elastic-
Transformation for brain tissue characterization
and segmentation," in Proceedings of IEEE
Western New York Image and Signal
Processing Workshop (WNYISPW), 2018, pp.
1-5.
[14] T. Yang, J. Song, and L. Li, "A deep learning
model integrating SK-TPCNN and random
forests for brain tumor segmentation in MRI,"
Biocybernetics and Biomedical Engineering,
2019, vol. 39, pp. 613-623.
[15] H. M. Ahmed, B. A. Youssef, A. S. Elkorany,
A. A. Saleeb, and F. Abd El-Samie, "Hybrid
gray wolf optimizer–artificial neural network
classification approach for magnetic resonance
brain images," Applied Optics, 2018, vol. 57,
pp. B25-B31.
[16] L. A. J. Prabhu and A. Jayachandran, "Mixture
model segmentation system for parasagittal
meningioma brain tumor classification based on
hybrid feature vector," Journal of medical
systems, 2018, vol. 42, pp. 251
Contribution of individual authors to
the creation of a scientific article
(ghostwriting policy)
Isselmou Abd El Kader: Created the Model, Writing,
Experiments simulation and analysis.
Guizhi Xu: Review& editing.
Zhang Shuai: Technical Review.
El maalouma Sidi Brahim: Writing Methodology.
Sani Saminu: Formatting.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.10
Isselmou Abd El Kader, Guizhi Xu,
Zhang Shuai, El Maalouma Sidi Brahim, Sani Saminu
E-ISSN: 2224-2902
83
Volume 19, 2022
Sources of funding for research
presented in a scientific article or
scientific article itself:
This work was supported by “the Natural Science
Foundation of China under Grant (51377045 and
Grant 31400844), This work was also supported by
the Specialized Research Fund for the Doctoral
Program of Higher Education under Grant
(20121317110002) and Grant (20131317120007).”
Creative Commons Attribution
License 4.0 (Attribution 4.0
International , CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2022.19.10
Isselmou Abd El Kader, Guizhi Xu,
Zhang Shuai, El Maalouma Sidi Brahim, Sani Saminu
E-ISSN: 2224-2902
84
Volume 19, 2022
