diagnostics

Article
Brain Tumor Detection and Classification on MR Images by a
Deep Wavelet Auto-Encoder Model
Isselmou Abd El Kader * , Guizhi Xu *, Zhang Shuai, Sani Saminu, Imran Javaid, Isah Salim Ahmad
and Souha Kamhi

State Key Laboratory of Reliability and Intelligence of Electrical Equipment, Hebei University of Technology,
Tianjin 300130, China; [email protected] (Z.S.); [email protected] (S.S.); [email protected] (I.J.);
[email protected] (I.S.A.); [email protected] (S.K.)
* Correspondence: [email protected] (I.A.E.K.); [email protected] (G.X.)

Abstract: The process of diagnosing brain tumors is very complicated for many reasons, including
the brain’s synaptic structure, size, and shape. Machine learning techniques are employed to
help doctors to detect brain tumor and support their decisions. In recent years, deep learning
techniques have made a great achievement in medical image analysis. This paper proposed a deep
wavelet autoencoder model named “DWAE model”, employed to divide input data slice as a tumor
(abnormal) or no tumor (normal). This article used a high pass filter to show the heterogeneity
of the MRI images and their integration with the input images. A high median filter was utilized
to merge slices. We improved the output slices’ quality through highlight edges and smoothened
input MR brain images. Then, we applied the seed growing method based on 4-connected since the





 thresholding cluster equal pixels with input MR data. The segmented MR image slices provide two
two-layer using the proposed deep wavelet auto-encoder model. We then used 200 hidden units
Citation: Abd El Kader, I.; Xu, G.;
Shuai, Z.; Saminu, S.; Javaid, I.;
in the first layer and 400 hidden units in the second layer. The softmax layer testing and training
Ahmad, I.S.; Kamhi, S. Brain Tumor are performed for the identification of the MR image normal and abnormal. The contribution of
Detection and Classification on MR the deep wavelet auto-encoder model is in the analysis of pixel pattern of MR brain image and the
Images by a Deep Wavelet ability to detect and classify the tumor with high accuracy, short time, and low loss validation. To
Auto-Encoder Model. Diagnostics train and test the overall performance of the proposed model, we utilized 2500 MR brain images
2021, 11, 1589. https://doi.org/ from BRATS2012, BRATS2013, BRATS2014, BRATS2015, 2015 challenge, and ISLES, which consists of
10.3390/diagnostics11091589 normal and abnormal images. The experiments results show that the proposed model achieved an
accuracy of 99.3%, loss validation of 0.1, low FPR and FNR values. This result demonstrates that the
Academic Editor: Panagiotis
proposed DWAE model can facilitate the automatic detection of brain tumors.
Papanagiotou

Keywords: MRI; brain tumor; detection; classification; seed growing; segmentation; deep wavelet
Received: 2 July 2021
auto-encoder
Accepted: 18 August 2021
Published: 31 August 2021

Publisher’s Note: MDPI stays neutral

1. Introduction
with regard to jurisdictional claims in
published maps and institutional affil- The brain tumor and its analysis are of extraordinary interest because of the devel-
iations. oping innovation in medical image processing. As indicated by the overview led by the
National Brain Tumor Foundation (NBTF), the improvement of brain tumor diagnosis
among patients and the death rate due to brain tumors is succeeding earlier year’s insights
across the globe [1,2]. The latest advances in machine learning (especially deep learning)
Copyright: © 2021 by the authors.
help identify, classify, and measure patterns in medical images. The core of these devel-
Licensee MDPI, Basel, Switzerland.
opments is by using hierarchical feature representations learned only from data rather
This article is an open access article
than manually designed features based on specific domain knowledge. Likewise, a few
distributed under the terms and works have proposed frameworks or models to feature the brain tumor zone in recent
conditions of the Creative Commons years, which might be trailed by stages such as outcome predictions, classification, and
Attribution (CC BY) license (https:// treatment planning. Brain tumor segmentation in medical image processing is necessary
creativecommons.org/licenses/by/ and generally governed by factors such as missing boundaries, noise, and low contrast.
4.0/). MRI segmentation using learning strategies and pattern recognition technology is very

Diagnostics 2021, 11, 1589. https://doi.org/10.3390/diagnostics11091589 https://www.mdpi.com/journal/diagnostics

Diagnostics 2021, 11, 1589 2 of 19

successful in analyzing brain images. Technically speaking, the method is a parametric

model that considers the functions selected based on the density function [3]. Early screen-
ing of such brain tumors issues is crucial to encourage convenient therapy and sound
living with contemporary clinical imaging modalities. The most common modalities that
are utilized to analyze the tumor in the brain are positron emission tomography (PET),
magnetic resonance imaging (MRI), and computed tomography (CT) [2].
Magnetic Resonance Imaging (MRI) is a well-known medical device used to diagnose
and analyze many diseases such as brain tumors, neurological diseases, epilepsy, etc.
Usually, a system completely processed by hardware/computer helps automate this process
to obtain accurate and fast results. On the other hand, image segmentation is the main
task of various computer vision and image processing implementations. The hypothesis
of the hash process divides the image into different areas according to some measures for
further processing [4,5]. Detection of brain abnormalities is usually done manually using
MRI imaging by medical experts. The large-scale manual examination method can often
lead to misinterpretation due to some factors such as fatigue and excessive abundance of
MRI slices. In addition, it is non-repeatable and results in intra- and inter-reader variability.
Alleviating these concerns requires developing a detection system method to diagnose
various brain abnormalities. It also helps in promoting fast, reliable, and accurate analysis
and supports the clinicians in their final selection process. Machine learning techniques
are mainly used to design and automate systems that have enjoyed spectacular success
in recent decades. Many methods (also known as automatic detection of pathological
brain systems) have been formulated to classify the brain’s different MRI scans. These
diagrams mainly focus on solving two brain types based on MRI classification disorders,
namely binary and multiclass. In the binary category, brain MRI scans are classified as
either pathological (abnormal) or normal.

2. Related Work
El-Dahshan et al. [6] proposed a method using two-dimensional deep wavelet trans-
form (2D-DWT) and principal component analysis (PCA) to extract salient features. They
employed a feed-forward neural network (FNN) and k-nearest neighbor (KNN) individ-
ually for classification. Das et al. [7] developed a model based on ripplet transform (RT),
and features are recharging to the least squares SVM (LS-SVM) classifier. A fluid vector is
utilized for tumor detection by using T1 weighted images [8]. Diffusion coefficients have
been used to identify the tumor based on the diffusion tensor images [9]. Researchers have
done a lot of work to extract and reduce an optimal feature of brain tumors; however, re-
moving and selecting the optimum feature remains a complicated task because the number
of features increases the association. In addition, selecting the training and testing samples
is also a challenge in obtaining good results [10,11]. Amin et al. [12] proposed a distinctive
method for MR brain detection and classification. First, the Gaussian filter was used to
eliminate noise; then, brain image features are extracted by embedded, cyclic, contrast,
block appearance, etc., for segmentation processing—cross-validation technology for classi-
fication. Chen et al. [13] suggested a hybrid between fuzzy clustering and Markov random
field and integrated the original image’s fuzzy clustering membership into Markov random
field function. This hybrid approach is the segmentation of the supporting information,
and it gives good efficiency.
Chen et al. [14] proposed a wavelet-like Auto Encoder (WAE) based on a neural
network, which analyzes the original image into a low-resolution image for classification.
These low-resolution channels or images are further used to input the Convolutional Neural
Network (CNN) to reduce computational complexity without altering the accuracy factor.
There are several deep convolutional neural networks and fully convolutional neural net-
works (FCNN) [15], two pathway cascaded neural network model [16], auto-encoder [17],
(CNN’s) [18,19], DeconvNets (EDD) [20], and three-dimensional convolutional neural net-
work (CNN) [21] are utilized for MRI images analysis. Two tracks CNN model [18] is used
for the prediction of brain tissues. Automatic multimodal brain tumor detection and classi-
Diagnostics 2021, 11, 1589 3 of 19

fication are discussed [22]. Binary CNN is used for complete tumor area prediction [23].
The patch-based approach is used for medical image analysis [24]. The victories of the deep
learning model were a game over lately, especially in 2012, Alexnet, the model that won the
Imagenet competition, was an important deciding opinion. The most significant difference
between a deep learning network and an artificial neural network is that a deep learning
network consists of several layers [25]. In addition, the capsule network is vastly used [26],
where the routing agreement process performs learning. The capsule network is excellent
for classification images compared to the CNN model. The pooling in capsule networks is
not performed for down sampling, as further improvement in tumor classification can also
be utilized for the fusion of deep learning and handcrafted features [27].
Korfiatis et al. [28] proposed a deep convolutional neural network method based on
autoencoders to segment increased signal areas in fluid-attenuated inversion recovery MRI
images. They trained automatic convolutional encoders on the BRATS Benchmark dataset
to segment the brain tumor image, and the accuracy was evaluated on a dataset with
three retail experts. They used the simultaneous truth and performance level estimation
(STAPLE) algorithm to provide the ground truth for comparison. The Jaccard coefficient,
dice coefficient, true-positive fraction, and false-negative fraction values were calculated.
The proposed method was within the inter-observer variability concerning true positive
fraction, Dice, and Jaccard. The developed approach can be utilized to output automatic
segmentation of the tumor area responsible for signal-increased fluid-attenuated inversion
recovery areas.
Kumar et al. [29] proposed a compression technique based on a deep wavelet autoen-
coder, which combines the fundamental feature reduction property of the auto encoder
and the image decomposition property of wavelet transform. These methods significantly
affect the feature set’s size for undertaking another classification task using DNN. A brain
image database was obtained, and the proposed DWA-DNN image classifier was consid-
ered. They compared the DWA-DNN classifier with the other classifiers such as DNN and
autoencoder and achieved better results. Deep Nayaka et al. [30] proposed a deep neural
network method where a stacked random vector functional link (RVFL) based autoencoder
(SRVFL-AE) is used to identify the different class brain abnormalities. The autoencoders
RVFL are the building blocks for their proposed SRVFL-AE. The main objective of selecting
RVFL as a critical component of the SRVFL-AE is to improve learning speed and general-
izability compared to deep learning methods based on the autoencoder. Moreover, they
incorporated a ReLU (Rectified Linear Unit) activation function into the deep network,
which they proposed for better-hidden representation of input features and better speed.
To assess the effectiveness of their approach, they took two standard datasets of MD-1 and
MD-2 MRI data. Their proposal approach achieved an accuracy of 96.67% and 95% on the
MD-1 and MD-2 datasets.
Mishra et al. [31] proposed an efficient method for magnetic resonance imaging
(MRI) brain image classification based on different wavelet transforms such as discrete
wavelet transform (DWT) and stationary wavelet transforms (SWT). Dual-tree M-band
wavelet transform (DMWT) was used for feature extraction and selection of coefficients
for classification using support vector machine classifiers. They decomposed the normal
and abnormal MRI brain image features through deep wavelet transform (DWT), SWT,
and DWT. The results of their proposed method achieved an accuracy of 98% for MR
brain images classification. Amin et al. [32] suggested a deep learning model to predict
input slices as a tumor (unhealthy)/non-tumor (healthy). They used a high pass filter
image to distinguish the MR slices’ in the homogeneities domain effect and integrated
them with the input slices. Then, they applied a median filter to fuse the slices. They
improved the quality of the resulting slices with smooth and highlighted edges of the input
slices. They segmented the slices to the fine-tuned two layers proposed stacked sparse
autoencoder (SSAE) approaches. They selected the hyperparametrs of the model after
extensive experiments. In the first layer, they used 200 hidden units, and on the second
layer, 400 hidden units. They tested the model on a softmax layer to predict from images
Diagnostics 2021, 11, 1589 4 of 19

with tumors and no tumors. They trained their model using BRATS datasets, i.e., 2012
(challenge and synthetic), 2013, and 2013 Leaderboard, 2014, and 2015 datasets.
Raja et al. [33] developed a brain tumor classification using a hybrid deep autoencoder
with a Bayesian fuzzy clustering-based segmentation method. They carried out a pre-
processing step with a non-local mean filter to reduce noise. They used the BFC (Bayesian
fuzzy clustering) method for the segmentation of brain tumors. After segmentation, they
used robust features such as information-theoretic measures, scattering transform (ST), and
wavelet packet Tsallis entropy (WPT) approaches for the feature extraction process. Finally,
they used a hybrid scheme of the DAE (deep autoencoder) based JOA (Jaya optimization
algorithm) with a softmax regression technique to classify the tumor area for the brain
tumor classification process. Their simulation result was conducted through the BRATS
2015 database.
Arunkumar et al. [34] proposed an improved automated brain tumor segmentation
and detection method using the ANN model. They used MR data without human medi-
ation by applying the best qualities for the preparatory detection of brain tumors. Their
brain tumor segmentation technique consists of three noteworthy improvement focuses.
Firstly, they used K-means clustering as part of the principal organization in MR data to
distinguish the areas of the district in light of their grayscale. Secondly, they used ANN to
select the correct object because of the training step. Thirdly, the tissue characteristics of
the brain tumor area were removed to the mitotic stage. In recognition of brain tumors,
gray-scale features analyze and diagnose brain tumors to distinguish benign and malig-
nant tumors. Their model evaluates and compares with the SVM method segmentation
outcomes and brain detection. Their model achieved an accuracy of 94.07%, sensitivity of
90.09%, and specificity of 96.78%.
Arunkumar et al. [35] suggested a novel segmentation study for brain tissues using
MR images. Their methods consist of three computer vision fiction steps: enhancing images,
segmenting images, and filtering out non-ROI based on the texture and HOG features. A
fully automated MRI-based brain tumor segmentation and classification method is based
on a model that uses artificial neural networks to locate an ROI accurately. Therefore,
the non-ROI filtering process was used for histogram examination to avoid non-ROI
and identify the correct object in brain MRI. However, histological features are used to
determine the type of tumor. Two hundred MRI cases were used to compare the automatic
and manual segmentation processes. The results show that fully automated trainable
model-based segmentation is superior to manual methods and brain recognition using
ROI texture features. Their model achieved a precision of 92.14%, with 89% sensitivity and
94% specificity.
Osama et al. [36] proposed a deep learning model that can predict mild cognitive
impairment (MCI), early MCI (EMCI), late MCI (LMCI), and AD Alzheimer’s disease
neuroimaging project. (ADNI) An fMRI data set consisting of 138 subjects was used for the
assessment. Their fine-tuned ResNet18 network model achieved an accuracy of 99.99%,
99.95%, and 99.95% on EMCI vs. AD, LMCI vs. AD, and MCI vs. EMCI classification
scenarios, respectively.
Huang et al. [37] proposed a differential feature map (DFM) block for brain tumor
detection using MR images. The DFM blocks are combined with stress and excitation (SE)
blocks to form a Differential Characteristic Neural Network (DFNN). First, they applied an
automatic image correction method to make the symmetry axis of the MRI image of the
brain approximately parallel to the vertical axis. In addition, DFNN was created to divide
brain MRI images into two categories: “abnormal” and “normal”. Their experimental
results show that the average accuracy of the proposed system on the two databases can
reach 99.2% and 98%, and the introduction of the proposed DFM block can increase the
average accuracy of the two databases by 1.8% and 1.3%, respectively.
Rundo et al. [38] suggested a fully automatic model for necrosis extraction (NeXt)
based on the Fuzzy C-Means algorithm after the GTV segmentation. Unsupervised machine
learning technology was used to detect and identify necrotic areas in heterogeneous
Diagnostics 2021, 11, 1589 5 of 19

cancers. Complete treatment pipeline is an integrated two-step segmentation method

that can be used to support neuroradiology. They used NeXt for dose escalation, allowing
more selective strategies to increase radiation dose in areas resistant to radiation and
hypoxia. In addition, NeXt only analyzes contrast-enhanced T1 MRI images and does not
require multispectral MRI data, which represents a clinically feasible solution. Their study
considers an MR database that consists of 32 metastatic brain cancers, wherein 20 tumors
present neuroses. The segmentation accuracy of the NeXt model was evaluated based on
spatial overlap-based and distance-based metrics values and achieved a dice similarity
coefficient of 95.93% ± 4.23% and mean absolute distance of 0.225 ± 0.229 (pixels).
Mekhmoukh et al. [39] proposed a novel segmentation approach based on Particle
Swarm Optimization (PSO) and outlier rejection combined with the level set. The tradi-
tional algorithm for brain tumor segmentation of the MR database is the fuzzy c-means
(FCM) algorithm. The membership function in this traditional algorithm is sensitive to
external factors and does not incorporate spatial information into the image. The algorithm
is very sensitive to noise and unevenness in the image, and it relies on the initialization
of the centroid. To improve the external suppression of traditional FCM aggregation algo-
rithms and reduce noise sensitivity, the authors presented a new extended FCM algorithm
for image segmentation. In general, in the FCM algorithm, the initial cluster centers are
selected randomly; with the help of the PSO algorithm, the centers of the clusters are
selected optimally. Their algorithm also takes into consideration spatial neighborhood
information. Their model achieved excellent effectiveness.
Han [40] proposed an unsupervised medical anomaly network model to detect un-
supervised medical anomalies (MADGAN). A new two-step method that uses GAN-
based multiple contiguous MRI slice reconstruction to detect brain abnormalities at dif-
ferent stages on multi-structured MRI: (reconstruction) Wasserstein loss with graded
penalty + 100, l1 loss − 1. Trained on three axial MRI slices of a healthy brain to rebuild
3-Reconstruct healthy/abnormal invisible scans; (diagnostics) average l2 loss − 2 per scan
characterizes them, by comparing baseline truth/reconstructed segments. They used two
different datasets for training: 1133 healthy T1-weighted (T1) and 135 healthy contrast-
enhanced T1 (T1c) brain MRI scans for detecting AD and brain metastases/various diseases,
respectively. Their MADGAN model can detect AD in very early T1 scans and mild cogni-
tive impairment (MCI), with an area under the curve (AUC) of 0.727 and advanced AD
with an AUC of 0.894. At the same time, it can detect brain metastases in T1c scans.
To overcome the current problems of deep learning approaches used in diagnosing
brain tumors, we still urgently need effective methods to diagnose brain tumors more
accurately in their early stages and more quickly to save time for doctors and increase the
patient’s survival rate [32,33]. Therefore, the proposed deep wavelet autoencoder (DWAE)
novelty is used to solve the problem of low performance, low loss validation, and longtime
processing while using MR brain big data analysis. The rest of the paper is organized as
follows: Section 1, introduction; Section 2, related work; the dataset collection is given in
Section 3; the methodology of the proposed model is given in Section 4; the experimental
results and discussion are given in Section 5; the conclusion and future works are given in
Section 6.

3. Datasets
In this work, we used five types of MR brain databases, including BRATS2012,
BRATS2013, BRATS2014, 2015 challenge, and Brats 2015, and ISLES. Figure 1 shows
different types of databases.
 Diagnostics
Diagnostics 11,11,
2021,
2021, 1589 PEER REVIEW
x FOR 66 of
of 19
20

(a)

(b)
Figure
Figure1. 1.
Sample (a) describe
(a) describe T1, T2-weighted,
T1, T2-weighted, and FLAIR
and FLAIR MR brain
MR brain images
images with tumor,
with tumor and
and (b) sample
describes
(b)T1,
describes T1, T2-weighted, and FLAIR MR brain images without
T2-weighted, and FLAIR MR brain images without the tumor. the tumor.

4.4.Methodology
Methodology
Theproposed
The proposedmodel
modelisisperformed
performedin inthree
threedifferent
different stages.
stages. InIn the
the first stage, a high
high
passfilter
pass filter[41]
[41]and
andthe
the median
median filter
filter [42]
[42] are
areselected
selectedtotoenhance
enhanceinput
inputMR MRbrain images.
brain In
images.
Inthe
thesecond
secondstage,
stage,the
theseed
seedgrowing
growing approach
approach [43][43] is
is utilized
utilized to
to segment
segment the the brain
brain tumor.
tumor.
Finally,ininthe
Finally, thethird
thirdstage,
stage,the
thesegmented
segmentedimages imagesarearesupplied
supplied toto the
the DWAE
DWAE model.
model. Every
Every
two hidden layers are being used, which are further connected with a softmax layer
two hidden layers are being used, which are further connected with a softmax layer for
for
classification.Figure
classification. Figure2 2shows
showsthe
theproposed
proposedapproach
approach stages.
stages.
Diagnostics 2021, 11, 1589 7 of 19
Diagnostics 2021, 11, x FOR PEER REVIEW 7 of 20

Figure 2. Proposed deep wavelet auto-encoder model architecture steps.

Figure 2. Proposed deep wavelet auto-encoder model architecture steps.

For more details, the main stages of the proposed approach are as follows:
1. The pre-processing stage is through an enhancement filter, to improve the image; we
introduce a new fusion method. In this step, the input MR brain images were resized
 Diagnostics 2021, 11, x FOR PEER REVIEW 8 of 20
Diagnostics2021,
Diagnostics 11,x1589
2021,11, FOR PEER REVIEW 8 8ofof20
19

by 256 × 256 × 1. Then, we choose a high pass filter to improve the edges of the input
by
For
MR256 × 256
more
brain × 1. Then,
details,
image. the we
Themain
inputchoose
and aoutput
stages ofhigh pass
the proposed
of the filter tobrain
improve
MRapproach arethe
image as
areedges
follows:
fusedofserially.
the input Fi-
1. MR brain
nally, a image.
combined, The input
fused MRand output
brain imageof the
is MR
smoothed brain
The pre-processing stage is through an enhancement filter, to improve the image; we image
using a are
3 × fused
3 medianserially.
filter Fi-
that
nally,
givesathe
introduce combined,
excellent
a new fusionfused
effectMR brainInimage
of segmentation
method. isresults
this step, smoothed
the input using
compared awith
MR brain 3 × images
3previous
median werefilter
models.that
resized
gives
2. by We256 the
applied excellent
× 256a × effect
seed-growingof segmentation
1. Then, wealgorithm
choose a based results
high pass compared
on the optimal
filter with previous
threshold
to improve models.
for good
the edges seg-
of the
2. We applied
mentation
input MR brain afor
seed-growing
aimage. The algorithm
brain tumor. input and outputbased on the MR
of the optimal
brainthreshold
image arefor good
fused seg-
serially.
3. mentation
Finally, a for
In classification, a brain
combined, wetumor.
applied
fused MRa brain
deep image
wavelet auto-encoder
is smoothed using(DWAE)
a 3 × 3model.
medianInfilter this
3. Inthat classification,
stage, the the
gives segmented we applied
excellent MR aofdeep
brain
effect image wavelet
is resized
segmentation auto-encoder
by 256
results × 256
compared (DWAE)
×with model.for
1 dimension
previous In faster
this
models.
2. stage,
We the segmented
processing.
applied aThe MR brain
objective
seed-growing image
ofalgorithm
this stage is based
resized onbythe
is to predict 256 × 256
the ×threshold
slices
optimal 1with
dimension for for
tumor faster
(abnormal
good seg-
processing.
MR brain images
mentation The objective
for a brainandtumor.of this stage is to predict the
the slices without tumor (normal MR brain images). slices with tumor (abnormal
3. MR brain images and
In classification, we the slicesawithout
applied deep wavelettumor auto-encoder
(normal MR brain (DWAE) images).
model. In this
4.1. Deep
stage,Wavelet MR brain image is resized by 256 × 256 × 1 dimension for faster
Auto-Encoder
the segmented
4.1. Deep
TheWavelet
normalAuto-Encoder
processing. The objective of
auto-encoder this stage
features is to predict
a strong inference theability,
slices with tumor (abnormal
robustness, and unsu-
MR
pervised brain
The normal images
featureauto-encoder and the
learning ability. slices
features without tumor
a strongofinference
The property (normal
the Wavelet MR
ability, brain images).
robustness,
transform andfeatures
has focal unsu-
pervised feature learning
and time-frequency ability. The
localization. property itofisthe
Therefore, Wavelettotransform
essential combine has focal features
standard auto-en-
4.1. time-frequency
and Deep Wavelet Auto-Encoder localization. Therefore, it is essential to combine standard auto-en-
coder and wavelet transform to solve the practical problems. This article proposed a new
coder The
and normal
wavelet auto-encoder
transform to features
solve thea strong
practical inference
problems.
kind of improved unsupervised neural network called the “deep wavelet auto-encoder” ability,
This robustness,
article and
proposed unsuper-
a new
vised
kind offeature
improved learning ability.
unsupervised The property
neural of
network the Wavelet
called
model, which can catch non-stationary vibration signals and represent complex infor- thetransform
“deep has
wavelet focal features
auto-encoder” and
time-frequency
model,
mation. which localization.
can catch
The wavelet Therefore,
non-stationary
auto-encoder model it is essential
vibration
utilizedsignals to combine standard
and represent
the wavelet function complex auto-encoder
infor-
as the activation
and
mation.wavelet
function The atransform
in wavelet to solve
auto-encoder
conventional state,thedefining
practical
model problems.
utilized
different the This article
wavelet
resolutions. proposed
function
The as the
wavelet aauto-encoder
new kind of
activation
improved
function
structure inisunsupervised
a shown
conventional neural
state,
in Figure network
defining
3, and called ofthe
different
the model “deep
resolutions.
the wavelet auto-encoder”
The wavelet
deep auto-encoder inmodel,
auto-encoder
is shown Figure
which
structure canis catch
shown non-stationary
in Figure 3, andvibration
the model signals
of theand deeprepresent complex
auto-encoder is information.
shown in The
Figure
4.
4.wavelet auto-encoder model utilized the wavelet function as the activation function in a
conventional state, defining different resolutions. The wavelet auto-encoder structure is
shown in Figure 3, and the model of the deep auto-encoder is shown in Figure 4.

Figure3.3.The
Figure Thestructure
structureof
ofthe
thewavelet
waveletauto-encoder.
auto-encoder.
Figure 3. The structure of the wavelet auto-encoder.

Figure4.4.AAmodel
Figure modelof
ofthe
thedeep
deepauto-encoder.
auto-encoder.
Figure 4. A model of the deep auto-encoder.
Diagnostics 2021, 11, 1589 9 of 19

Figure 4 represents the working principle of the deep auto-encoder, and Equation (1)
indicates the decoding stages.
X = ξ (κ̂ 0 Y + b0 ) (1)
where X̂ indicates the outcome of the reconstructed vector, k represents kernel vector, b0
denotes bias value, and ∈ is an error value added during backpropagation.

Deep Wavelet Auto-Encoder Model Training

Training samples y = [y1 , y2 , . . . , yn ] A the output of the hidden unit is i.

∑in−1 vij yl − ei


gi (out) = ϕ (2)
bi

where:
ϕ Represents the wavelet activation function.
yl (r = 1, 2, . . . , n) Is the lth dimension input of the training sample?
vij (i = 1, 2, 3, . . . , g) Is the connection weight between input unit l and the hidden
unit i.
bi and ei represent vij (i = 1, 2, 3, . . . , g) sent the scale factor and shift factor of wavelet
activation function for a hidden unit i.
 
ϕ( a) = cos(5a) exp a2 /2 (3)

2 n 2 !
∑nl−1 vil yl − ei 1 ∑l −1 vij yl − ei

gi (out) = ϕb.e (i ) = cos 5 × × − (4)
bi 2 bi

Similar to a normal auto-encoder (AE), we choose the output layer’s activation func-
tion as sigmoid function. T hen, the output of the deep wavelet auto-encoder can be
calculated as in Equation (5):
 
!2  
n n
∑ ∑


q v y − e − 1 v y − e
ŷ = sigm∑i−1 vri cos 5 × i −1 il l i l −1 il l i
× exp  (5)
bi 2 bi

where ŷ is i the reconstructed dimension output of the training samples, and vri is the
connection weight between hidden r and i?

4.2. High Pass Filter

Due to the fluctuations of the magnetic field, noise appears in the MR image acquisition
process. Therefore, an approach was proposed for enhancing the lesion region and remove
the noise. In the proposed method, the high pass filter chooses to distinguish the edges in
the input I. We obtained the sharper image shI by combining I with the filtered image hpI
as in Equation (6):
sh I ( x,y) = I ( x, y) + hp I ( x,y) (6)
To smooth the sharper image’s intensities, we utilized the median filter with a 3 × 3
window size.

4.3. Segmentation Using a Seeded Region Growing

By using a seed growing method, brain tumor segmentation is performed. The iden-
tical pixel values are grouped when different iterations are completed. We repeated this
process until the area stops growing. Equation (7) mathematically describes the predicated
n function prI utilized in the seed-growing algorithm.
1
pr I ( x,y) = 0 If the difference between the seed point and current pixels ≤ T otherwise (7)

T Denotes the predefined threshold.

Diagnostics 2021, 11, 1589 10 of 19

4.4. Softmax Classifier

This paper chooses the softmax to connect the deep wavelet network as follows in
Equation (8).
ek Y
ψ(Yi ) = N i i (8)
∑i− j ek j Yi
where Yi indicates the output of DWAE.
The softmax classifier ψ is utilized for classification based on probability. ki Denotes
ith neuron until kernel vector and N shows total classes. The output ψ(Yi ) shows the ith
class probability.
The segmented MR brain images are provided for training and testing in the DWAE
model. The input image is resized to 32 × 32 × 500, where 500 corresponds to the number
of the slices. The input image 256 × 256 describes that 1024 units are provided to the DWAE
model. The input with 200 hidden units is passed to DWAE1. DWAE1 with weight initial
feature vector and bias is produced, which is given as input to the DWAE2. In DWAE2 400,
hidden units are utilized with weight, preference, and hyperparametrs to the produced
feature vector. We utilized the softmax layer to perform the classification based on the
feature vector acquired from DWAE2.

5. Experimental Results and Discussion

In this work, the deep wavelet auto-encoder model is used for training and testing
using different databases, as shown in the dataset section. Database 2012 consists of 500 MR
brain images in the training stage and 500 MR images, including abnormal and normal
in the testing stages [43]. The BRATS2013 database consisting of 1000 MR brain images
is used [15]. The BRATS2014 database consists of 800 MR brain Images [44]. The 2015
challenge and Brats 2015, ISLES consists of 700 MR brain images [45]. The input slices
consist of abnormal and normal MR brain images from the datasets above. We ran this
simulation using a ThinkStation P620 Tower Workstation, NVIDIA Quadro® P2200 16 GB,
Lenovo Company, Tianjin City, China. To create a DWAE architecture, we used Tensorflow
and Keras, Spyder 3.6. We tested different databases to evaluate and analyze different
factors in the deep wavelet auto-encoder model. The results of the experiments include two
steps. Firstly, the results of MR database processing, and secondly, performance metrics.

5.1. The Results of MR Database Processing

This section used the deep wavelet auto-encoder model to process MR dataset and
the segmentation method described in Figure 5 and the classification method, as shown in
Figures 6 and 7.
Figure 5 represents the results of pre-processing steps based on the segmentation
method, which consists of the (a) original image, (b) the result of applied image sharpening,
(c) the results of applied high pass filter, (d) the result of applied seed growing, (e) the
result of applied thresholding and (f) the result of the applied segment tumor area. Figure 5
proves the excellent ability of the proposed DWAE model during pre-processing stage
and segmentation.
Figure 6 describes the classification stage and shows the excellent capacity of the
proposed model in the classification stage using normal MR databases. The slice consists of
15 MR normal brain images processed using a classification method based on the DWAE
model, and the output shows a good result. Figure 7 represents the classification stage,
and notable is the perfect ability of the proposed model in the classification stage using an
abnormal database. The slice consists of 15 MR abnormal brain images processed using the
classification method based on the proposed model. The results show the robustness of
the model to classify the brain tumor. Figures 6 and 7 prove the ability of the proposed
model to predict the normal MR brain images and abnormal MR brain images. Generally,
the model demonstrated a high ability to process, detect, and classify brain tumors using
MR images.
 Diagnostics 2021, 11, 1589 11 of 19
Diagnostics 2021, 11, x FOR PEER REVIEW 11 of 20

(a) (b) (c) (d) (e) (f)

Figure 5. (a) Original image, (b) image sharpening, (c) high pass filter, (d) seed growing, (e) applied
thresholding, and (f) segmented tumor alone.

Figure 5 represents the results of pre-processing steps based on the segmentation

method, which consists of the (a) original image, (b) the result of applied image sharpen-
ing, (a)
(c) the results (b)of applied high
(c) pass filter,(d)(d) the result(e)of applied seed(f) growing, (e)
the result of applied thresholding and (f) the result of the applied segment tumor area.
Figure 5. (a) Original image, (b)
(b)image sharpening, (c)
(c) high pass filter,
filter, (d)
(d)seed
seedgrowing,
growing,(e)
(e)applied
applied
Figure5.5(a)
Figure Original
proves theimage,
excellentimage sharpening,
ability high pass
of the proposed DWAE model during pre-processing
thresholding,
thresholding, and (f) segmented tumor alone.
and (f) segmented tumor alone.
stage and segmentation.
Figure 5 represents the results of pre-processing steps based on the segmentation
method, which consists of the (a) original image, (b) the result of applied image sharpen-
ing, (c) the results of applied high pass filter, (d) the result of applied seed growing, (e)
the result of applied thresholding and (f) the result of the applied segment tumor area.
Figure 5 proves the excellent ability of the proposed DWAE model during pre-processing
stage and segmentation.

Figure6.6. Samples
Figure Samplesshowing
showingthe
theresults
resultsofofMR
MRbrain
brainnormal
normalimages
images classification
classification using
using thethe deep
deep
wavelet auto-encoder model.
wavelet auto-encoder model.

The Figure 8 represents the accuracy value of DWAE model (training and testing), the
graph shows the effectiveness of the proposal model by achieved accuracy of 99.3%.
The loss value (training and testing) means that validation loss has the same metric
as training loss, but it is not utilized to update the weights. It is calculated similarly by
Figure 6. Samples showing the results of MR brain normal images classification using the deep
running the network forward over inputs xi and comparing the network outputs ŷi with
wavelet auto-encoder model.
the ground truth-values using a loss function (9).

1 N
N ∑ i −1
J= ζ (ŷi , yi ) (9)
Diagnostics 2021, 11, 1589 12 of 19

Figure 9 presents the proposed DWAE model loss value; the results show that the
proposed model achieved (0.1 to 0.3) a best loss validation value during the training and
Diagnostics 2021, 11, x FOR PEER REVIEW
testing stages. We trained and tested the proposed model using 30 epochs during12the of 20

accuracy and loss validation stages.

Figure 7. Samples showing the results of MR brain abnormal images classification using
wavelet auto-encoder model.

Figure 6 describes the classification stage and shows the excellent capacity o
posed model in the classification stage using normal MR databases. The slice co
15 MR normal brain images processed using a classification method based on th
model, and the output shows a good result. Figure 7 represents the classificati
and notable is the perfect ability of the proposed model in the classification stage
abnormal database. The slice consists of 15 MR abnormal brain images process
the classification method based on the proposed model. The results show the ro
of the model to classify the brain tumor. Figures 6 and 7 prove the ability of the p
model to predict the normal MR brain images and abnormal MR brain images. G
the model demonstrated a high ability to process, detect, and classify brain tum
MR images.
The Figure 8 represents the accuracy value of DWAE model (training and
Figure7.7.Samples
Figure Samplesshowing
showingthe
theresults
resultsofofMR
MR brain
brain abnormal
abnormal images
images classification
classification using
using thethe deep
deep
the graph shows the effectiveness
wavelet auto-encoder model.
of the proposal model by achieved accuracy o
wavelet auto-encoder model.

Figure 6 describes the classification stage and shows the excellent capacity of the pro-
posed model in the classification stage using normal MR databases. The slice consists of
15 MR normal brain images processed using a classification method based on the DWAE
model, and the output shows a good result. Figure 7 represents the classification stage,
and notable is the perfect ability of the proposed model in the classification stage using an
abnormal database. The slice consists of 15 MR abnormal brain images processed using
the classification method based on the proposed model. The results show the robustness
of the model to classify the brain tumor. Figures 6 and 7 prove the ability of the proposed
model to predict the normal MR brain images and abnormal MR brain images. Generally,
the model demonstrated a high ability to process, detect, and classify brain tumors using
MR images.
The Figure 8 represents the accuracy value of DWAE model (training and testing),
the graph shows the effectiveness of the proposal model by achieved accuracy of 99.3%.

Figure 8. The result of the proposed model accuracy value and the model’s accurate and robust
Figure 8. The result of the proposed model accuracy value and the model’s accurate a
model during training and testing stages as calculated using Equation (10).
model during training and testing stages as calculated using Equation (10).
5.2. Performances Metrics
To evaluate the proposed DWAE model efficiency, we compare the proposed model
overall performance using accuracy, sensitivity, specificity, DSC, Precision, JSI, FPR, and
FNR values.

Figure 8. The result of the proposed model accuracy value and the model’s accurate and robust
model during training and testing stages as calculated using Equation (10).
 1
  yi , yi 
N
J
N
 i 1
 
(9)

Figure 9 presents the proposed DWAE model loss value; the results show that the
proposed model achieved (0.1 to 0.3) a best loss validation value during the training and
Diagnostics 2021, 11, 1589
testing stages. We trained and tested the proposed model using 30 epochs during 13 of 19
the ac-
curacy and loss validation stages.

Describethe
Figure9.9.Describe
Figure theresults
results of
of the
the loss
loss value
value using
using the
the DWAE
DWAEmodel.
model.
5.2.1. Accuracy (ACC)
5.2. Performances Metrics
Accuracy (ACC) is utilized to compute the degree of correct tumor classification rate,
and To evaluate the
is calculated proposed
using DWAEEquation
the following model efficiency,
(10): we compare the proposed model
overall performance using accuracy, sensitivity, specificity, DSC, Precision, JSI, FPR, and
FNR values. (TP + TN)
Accuracy = × 100 (10)
(TP + TN) + (FP + FN)
5.2.1. Accuracy (ACC)
5.2.2. Sensitivity (SE)
Accuracy (ACC) is utilized to compute the degree of correct tumor classification rate,
and isSensitivity
calculated(SE) is utilized
using to calculate
the following the degree
Equation (10): of how much approach is sensitive to
measure the tumor identification rate, and is calculated using the following Equation (11):
TP  TN
Accuracy = (TP) × 100
 100
 (10)
Sensitivity (11)
(TP
TP+FN
TN 
)  ( FP  FN )

5.2.3. Specificity (SP)

5.2.2. Sensitivity (SE)
Specificity (SP) is the rate between true negative (TN) and true positive (TP), and is
Sensitivity
calculated using(SE)
the is utilized Equation
following to calculate the degree of how much approach is sensitive
(12):
to measure the tumor identification rate, and is calculated using the following Equation
(11): (TN)
Specificity = × 100 (12)
(TN + FP)

Sensitivity 
TP   100
5.2.4. Dice Similarity Coefficient (DSC) (11)
TP  FN 
Dice similarity coefficient (DSC) is utilized to compute the ratio between the actual
tumor and non-tumor, which are compared with predicted tumor and non-tumor pixels,
and is calculated using the following Equation (13):

2TP
DSC = × 100 (13)
FP + 2TP + FN

5.2.5. PRECISION (PRE)

PRECISION (PRE) describes the number of digits that are used to express a value, and
is calculated using the following Equation (14):

(TP)
Precision = × 100 (14)
(TP + FP)
Diagnostics 2021, 11, 1589 14 of 19

5.2.6. JACCARD Similarity Index (JSI)

JACCARD similarity index (JSI) is utilized to compute the similarity between the
actual tumor pixels and predicted tumor pixels and is calculated using the following
Equation (15):
TP
JSI = × 100 (15)
TP + FN + FP

5.2.7. FALSE Positive Rate (FPR)

FALSE positive rate (FPR) is utilized to compute the ratio of wrongly identified pixels,
corrected identified pixels, and is calculated using the following Equation (16):

FPR = 1 − Specificity (16)

5.2.8. FALSE Negative Rate (FNR)

FALSE negative rate (FNR) is utilized to compute the positive proportion, but the
approach-identified negative and is calculated using the following Equation (17).

FNR = 1 − Sensitivity (17)

where true positive (TP), true negative (TN), false positive (FP), and false negative (FN).
The proposed DWAE model is evaluated by comparing 21 existing models on machine
learning methods based on the accuracy (ACC), sensitivity (SE), specificity (SP), dice
similarity coefficient (DSC), precision (PR), Jaccard similarity index (JSI), false-positive rate
(FPR), and false-negative rate (FNR) values as shown in the Tables 1–4.

Table 1. Comparison of the results of previous models with the proposed DWAE model.

Model Accuracy % Sensitivity % Specificity % Precision %

DWA-DNN [29] 93.14 92.16 94.26 93.15
DAE-JOA [33] 98.5 95.4 - 95.6
CNN [46] 96.5 - 95 94.81
Google-Net [47] 89.66 84.85 96 96.55
Vgg16 [47] 84.48 81.25 88.48 89.66
KNN [48] 78 46 50 52
DNN [49] 93 75 80 72
M-CNN [50] 96.4 95 93 95.7
CNN-SVM [51] 95.62 - 95 92.12
Alex-Net [52] 87.66 84.38 92.31 93.1
HOG + LBP + Deep
98.71 98.46 96.72 -
features [53]
RG + MKM + U-NET [54] 98.72 90.7 99.7 -
Proposed DWAE Model 99.3 95.6 96.9 97.4

Table 1 compared the proposed DWAE model results with 10 previous models pub-
lished in the best level indexed. Deep wavelet autoencoder with deep neural network
DWA-DNN model [29] for brain MRI image classification for cancer identification achieved
an accuracy of 93.14%, sensitivity of 92.16%, and specificity of 94.26%, and precision
of 94.81%.
Brain tumor detection using the convolutional neural network (CNN) model [46]
obtained an accuracy of 96.5%, specificity of 95%, and a precision of 89.66%. VGG16 model
for brain classification and analysis [47] achieved an accuracy of 84.48%, sensitivity of
81.25%, specificity of 88.48%, and precision of 89.66%.
Diagnostics 2021, 11, 1589 15 of 19

Brain tumor classification using the k-nearest neighbors (KNN) model [48] obtained
an accuracy of 78%, a sensitivity of 46%, and a specificity of 50%. The deep neural network
(DNN) model for brain cancer detection [49] achieved an accuracy of 93%, sensitivity of
75%, specificity of 80%, and precision of 72%.
Brain tumor classification using modified convolutional neural network (M-CNN)
model [50] obtained an accuracy of 96.4%, a sensitivity of 95%, a specificity of 93%, and a
precision of 95.7%. Brain tumor segmentation using deep auto-encoder with Jaya optimiza-
tion algorithm (DAE-JOA) model [33] achieved an accuracy of 98.5%, sensitivity of 95.4%,
and precision of 95.6%.
A hybrid convolutional neural network with a self-vector machine (CNN-SVM)
model [51] for brain tumor classification obtained an accuracy of 95.62%, specificity of
95%, and precision of 92.12%. Brain tumor identification using the Alex-Net model [52]
achieved an accuracy of 87.66%, sensitivity of 84.38%, and specificity of 92.31%. Google-Net
model [47] for brain tumor detection and analysis obtained an accuracy of 89.66%, a sensi-
tivity of 84.85, specificity of 96%, and precision of 97.4%. The deep features model [53] for
brain tumor achieved an accuracy of 98.71% and specificity of 96.71%. U-NET model [54]
for brain tumor detection obtained an accuracy of 98.72% and sensitivity of 90.7%. The pro-
posed deep wavelet auto-encoder (DWAE) model obtained an accuracy of 99.3%, sensitivity
of 95.6%, specificity of 96.9% and precision of 97.4%.
Based on the results of 10 existing models shown in Table 1, we conclude that the
proposed deep wavelet auto-encoder (DWAE) model achieved better performance than
previous models using accuracy, sensitivity, specificity, and precision values.
Table 2 describes a comparison between the proposed deep wavelet auto-encoder
(DWAE) models results with existing models are based on dice similarity coefficient (DSC)
value achievement. Brain tumor classification using the convolutional neural network
(CNN) model [20] achieved a DSC of 83.7%. The convolutional neural network small filter
model [49] obtained a DSC of 88% for brain tumor classification. Brain tumor detection
using the conditional random fields (CRF) model achieved a DSC of 62%.

Table 2. Comparison of the existing model with the proposed DWAE model using DSC value.

Model DSC%
CNN [20] 83.7
CNN-small filter [49] 88
CRF [55] 62
HMV [45] 85
3D fully connected [21] 84.7
Integrated hierarchical [56] 73
Local independent projection [57] 75
RG + MKM + U-NET [54] 90
HOG + LBP + deep features [53] 96.11
Multi-scale 3D with fully connected CRF [21] 90
Proposed DWAE model 96.55

The hierarchical majority vote (HMV) model for brain tumor detection [45] obtained a
DSC of 85%. Brain tumor classification using a 3D fully connected model achieved a DSC
of 84.7%. The integrated hierarchical model [47] for brain tumor detection obtained a DSC
of 73%.
Local independent projection model [57] for brain tumor classification achieved a
DSC of 75%. Brain tumor classification using multi-scale 3D with fully connected CRF
model [21] achieved DSC of 90%. The deep features model [53] for brain tumor achieved a
Diagnostics 2021, 11, 1589 16 of 19

DSC of 90%. U-NET model [54] for brain tumor detection obtained a DSC of 96.1%. The
proposed DWAE model obtained a DSC of 96.55%.
According to the results shown in Table 2, based on the nine previous models, we
noticed that the proposed deep wavelet auto-encoder (DWAE) model achieved a better
dice similarity coefficient (DSC) value than the existing models.
Table 3 presented the comparison results of the proposed DWAE model with existing
models using FPR and FNR values. Brain tumor detection using the deep neural network
(DNN) model [58] achieved FPR 0.16 and FNR 0.06. The deep autoencoder with Jaya opti-
mization algorithm (DAE-JOA) model for brain tumor detection [33] obtained FPR 0.46 and
FNR 0.04. Stacked auto-encoder model for brain tumor identification [32] achieved FPR 0.07
and FNR 0.1. Brain tumor classification based on the Alex-Net model [52] obtained FPR 0.07
and FNR 0.339. Google-Net model [47] for brain tumor detection achieved FPR 0.714 and
FNR 0.339. Brain tumor identification using k-nearest neighbors (KNN) model [48] ob-
tained FPR 0.62 and FNR 0.54. The multimodal model for brain tumor classification using
deep learning model [59] achieved an FNR of 1.74. The proposed DWAE model achieved
FPR 0.0625 and FNR 0.031. Based on the results shown in Table 3, the proposed model
produced better results using FPR and FNR values than the previous models.

Table 3. Comparison of the existing models with the proposed DWAE model using FPR and
FNR values.

Model FPR FNR

DNN [58] 0.16 0.06
DAE-JOA [33] 0.46 0.04
Stacked auto-encoder [32] 0.07 0.1
Alex-Net [52] 0.07 0.128
Google-Net [47] 0.714 0.339
Multimodal [59] - 1.74
KNN [48] 0.62 0.54
Proposed DWAE model 0.0625 0.031

Table 4 describes a comparison of the proposed DWAE model’s results with an existing
model based on the JSI value. The hybrid stacked auto-encoder based on deep learning
model [32] for brain tumor detection obtained JSI of 89%. A deep neural network model [58]
for MR big data analysis achieved a JSI of 90.4%. A stable algorithm based on a deep
learning model [28] for automated segmentation using MR FLAIR images obtained a JSI of
92.3%. Our proposed model achieved a JSI of 93.3%.

Table 4. Comparison of the previous model with the proposed DWAE model using JSI value.

Model JSI%
Stacked auto-encoder [32] 89
DNN [52] 90.4
Stable algorithm [28] 92.3
Proposed DWAE model 93.3

The proposed DWAE model is compared with 21 previous models such

as [20,21,28,29,32,33,45–59], as shown in Tables 1–4 based on accuracy, sensitivity, speci-
ficity, precision, DSC, FPR, FNR, and JSI values. According to the performance and analysis
of the previous models above, the results suggest that the proposed model performed very
well compared to previous models.
In this work, we discuss threats-to-validity of the experimental results using 21 exist-
ing models published in the top-level journals in the field of brain tumor detection and
classification using the same databases as shown in Tables 1–4. The validation of our
results is based on a comparison between existing models with the proposed deep wavelet
Diagnostics 2021, 11, 1589 17 of 19

auto-encoder model using eight values. The ability of the proposed model to be validated
on other BRATS big databases with high accuracy and very low loss validation, as shown
in Figures 8 and 9, show a clear evidence of the threat-validity of the results.

6. Conclusions
Deep learning networks models have obtained good results in recent years in the
medical image analysis field. In this model, we implemented the necessary phases such as
image sharpening, high pass filter, thresholding segmentation, growing seed approach, and
classification based on deep wavelet auto-encoder model for feature extraction; the imple-
mentation produced excellent results, as shown in the results experiments in Figures 5–7.
The proposed model testing and training combine databases from BRATS2012, BRAT-
S2013, BRATS2014, 2015 challenge, and Brats 2015. The average accuracy is 99.3%, sensitiv-
ity 95.6%, specificity 96.9%, precision 97.4%, DSC 96.55%, FPR 0.0625, FNR 0.031, and JSI
93.3%, respectively. Based on the overall experiment’s output, segmentation, classification,
and performance of the proposed DWAE model, we conclude that the proposed model
achieved better results than the 21 existing models published in high-level journals.
The advantage of this proposed method is its excellent ability to analyze large data
from magnetic resonance images of the brain without technical problems and with very
high accuracy, which will help doctors in the accurate diagnosis of brain tumors.
Our proposed model achieved a great overall performance on brain tumor identifi-
cation and classification stages, allowing the model to be used in computing techniques
for the early detection of brain tumor. DWAE model shows the importance of the deep
learning model in the medical field and medical applications. In future work, we will
evaluate the overall performance of our DWAE model. We will improve the different layer
parameters that are hidden in the model to increase the accuracy and make the model faster.
We will validate our proposed model using the BRATS 2019 and BRATS 2020 datasets.

Author Contributions: I.A.E.K.: writing, simulation, and analysis. G.X.: review and editing. Z.S.:
a technical review. S.S.: language review. I.J.: writing methodology and I.S.A.: formatting: S.K.:
formatting. All authors have read and agreed to the published version of the manuscript.
Funding: 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).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data presented in this study are available on request from the
corresponding author.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. El-Dahshan, E.-S.A.; Mohsen, H.M.; Revett, K.; Salem, A.-B.M. Computer-aided diagnosis of human brain tumor through MRI: A
survey and a new algorithm. Exp. Syst. Appl. 2014, 41, 5526–5545. [CrossRef]
2. Logeswari, T.; Karnan, M. An improved implementation of brain tumor detection using segmentation based on the hierarchicaself-
organizing map. Int. J. Comp. Theory Eng. 2010, 2, 591. [CrossRef]
3. Tong, J.; Zhao, Y.; Zhang, P.; Chen, L.; Jiang, L. MRI brain tumor segmentation based on texture features and kernel sparse coding.
Biomed. Signal Process. Control 2018, 47, 387–392. [CrossRef]
4. Singh, M.; Verma, A.; Sharma, N. An Optimized Cascaded Stochastic Resonance for the Enhancement of Brain MRI. Innov. Res.
Biomed. Eng. J. 2018, 39, 334–342. [CrossRef]
5. Pham, T.X.; Siarry, P.; Oulhadj, H. Integrating fuzzy entropy clustering with an improved PSO for MRI brain image segmentation.
Appl. Soft Comput. 2018, 65, 230–242. [CrossRef]
6. El-Dahshan, E.-S.A.; Hosny, T.; Salem, A.-B.M. Hybrid intelligent techniques for MRI brain images classification. Digit. Signal
Process. 2010, 20, 433–441. [CrossRef]
7. Das, S.; Chowdhury, M.; Kundu, M.K. Brain MR image classification using multiscale geometric analysis of ripplet. Prog.
Electromagn. Res. 2013, 137, 1–17. [CrossRef]
Diagnostics 2021, 11, 1589 18 of 19

8. Wang, T.; Cheng, I.; Basu, A. Fluid Vector Flow and Applications in Brain Tumor Segmentation. IEEE Trans. Biomed. Eng. 2009, 56,
781–789. [CrossRef] [PubMed]
9. Al-Okaili, R.N.; Krejza, J.; Woo, J.H.; Wolf, R.L.; O’Rourke, D.M.; Judy, K.D.; Poptani, H.; Melhem, E.R. Intraaxial brain masses:
MR imag-ing-based diagnostic strategy—Initial experience. Radiology 2007, 243, 539–550. [CrossRef] [PubMed]
10. Jiang, J.; Wu, Y.; Huang, M.; Yang, W.; Chen, W.; Feng, Q. 3D brain tumor segmentation in multimodal MR images based on
learning population- and patient-specific feature sets. Comput. Med. Imaging Graph. 2013, 37, 512–521. [CrossRef]
11. Ortiz, A.; Górriz, J.M.; Ramírez, J.; Salas-Gonzalez, D. Improving MRI segmentation with probabilistic GHSOM and multiobjective
optimization. Neurocomputing 2013, 114, 118–131. [CrossRef]
12. Amin, J.; Sharif, M.; Yasmin, M.; Fernandes, S.L. A distinctive approach in brain tumor detection and classification using MRI.
Pattern Recognit. Lett. 2020, 139, 118–127. [CrossRef]
13. Chen, M.; Yan, Q.; Qin, M. A segmentation of brain MRI images utilizing intensity and contextual information by Markov random
field. Comput. Assist. Surg. 2017, 22, 200–211. [CrossRef]
14. Chen, T.; Lin, L.; Zuo, W.; Luo, X.; Zhang, L. Learning a wavelet-like auto-encoder to accelerate deep neural networks. In
Proceedings of the AAAI Conference on Artificial Intelligence, New Orleans, LA, USA, 2–7 February 2018; pp. 74–93.
15. Zhao, X.; Wu, Y.; Song, G.; Li, Z.; Zhang, Y.; Fan, Y. A deep learning model integrating FCNNs and CRFs for brain tumor
segmentation. Med. Image Anal. 2018, 43, 98–111. [CrossRef] [PubMed]
16. Havaei, M.; Dutil, F.; Pal, C.; Larochelle, H.; Jodoin, P.-M. A Convolutional Neural Network Approach to Brain Tumor Segmenta-
tion. In BrainLes; Springer: Cham, Switzerland, 2015; pp. 195–208.
17. Chen, X.; Konukoglu, E. Unsupervised detection of lesions in brain MRI using constrained adversarial auto-encoders. arXiv 2018,
arXiv:1806.04972.
18. Havaei, M.; Davy, A.; Warde-Farley, D.; Biard, A.; Courville, A.; Bengio, Y.; Pal, C.; Jodoin, P.-M.; Larochelle, H. Brain tumor
segmentation with Deep Neural Networks. Med. Image Anal. 2017, 35, 18–31. [CrossRef] [PubMed]
19. Dvořák, P.; Menze, B. Local Structure Prediction with Convolutional Neural Networks for Multimodal Brain Tumor Segmentation.
In International MICCAI Workshop on Medical Computer Vision; Springer: Cham, Switzerland, 2015; pp. 59–71.
20. Chen, L.; Bentley, P.; Rueckert, D. Fully automatic acute ischemic lesion segmentation in DWI using convolutional neural
networks. NeuroImage Clin. 2017, 15, 633–643. [CrossRef]
21. Kamnitsas, K.; Ledig, C.; Newcombe, V.; Simpson, J.P.; Kane, A.D.; Menon, D.K.; Rueckert, D.; Glocker, B. Efficient multi-scale 3D
CNN with fully connected CRF for accurate brain lesion segmentation. Med. Image Anal. 2017, 36, 61–78. [CrossRef]
22. El-Melegy, M.; El-Magd, K.M.A.; Ali, S.A.; Hussain, K.F.; Mahdy, Y.B. A comparative study of classification methods for automatic
multimodal brain tumor segmentation. In Proceedings of the International Conference on Innovative Trends in Computer
Engineering (ITCE), Aswan, Egypt, 19–21 February 2018; pp. 36–41.
23. Lyksborg, M.; Puonti, O.; Agn, M.; Larsen, R. An Ensemble of 2D Convolutional Neural Networks for Tumor Segmentation. In
19th Scandinavian Conference on Image Analysis; Springer: Copenhagen, Denmark, 2015; pp. 201–211.
24. Rao, V.; Sarabi, M.S.; Jaiswal, A. Brain tumor segmentation with deep learning. In Proceedings of MICCAI Multimodal Brain Tumor
Segmentation Challenge (BraTS); Springer: Cham, Switzerland, 2015; Volume 59, p. 56.
25. Cengil, E.; Çınar, A.; Özbay, E. Image classification with caffe deep learning framework. In Proceedings of the International
Conference on Computer Science and Engineering (UBMK), Antalya, Turkey, 5–8 October 2017; pp. 440–444.
26. Afshar, P.; Mohammadi, A.; Plataniotis, K.N. Brain Tumor Type Classification via Capsule Networks. In Proceedings of the 25th
IEEE International Conference on Image Processing (ICIP), Athens, Greece, 7–10 October 2018; pp. 3129–3133.
27. Akkus, Z.; Galimzianova, A.; Hoogi, A.; Rubin, D.L.; Erickson, B.J. Deep Learning for Brain MRI Segmentation: State of the Art
and Future Directions. J. Digit. Imaging 2017, 30, 449–459. [CrossRef]
28. Korfiatis, P.; Kline, T.L.; Erickson, B.J. Automated Segmentation of Hyper intense Regions in FLAIR MRI Using Deep Learning.
Tomography 2016, 2, 334–340. [CrossRef] [PubMed]
29. Mallick, P.K.; Ryu, S.H.; Satapathy, S.K.; Mishra, S.; Nguyen, G.N.; Tiwari, P. Brain MRI Image Classification for Cancer Detection
Using Deep Wavelet Autoencoder-Based Deep Neural Network. IEEE Access 2019, 7, 46278–46287. [CrossRef]
30. Nayak, D.R.; Dash, R.; Majhi, B.; Pachori, R.B.; Zhang, Y. A deep stacked random vector functional link network autoencoder for
diagnosis of brain abnormalities and breast cancer. Biomed. Signal Process. Control 2020, 58, 101860. [CrossRef]
31. Mishra, S.K.; Deepthi, V.H. Brain image classification by the combination of different wavelet transforms and support vector
machine classification. J. Ambient. Intell. Humaniz. Comput. 2020, 12, 6741–6749. [CrossRef]
32. Amin, J.; Sharif, M.; Gul, N.; Raza, M.; Anjum, M.A.; Nisar, M.W.; Bukhari, S.A.C. Brain Tumor Detection by Using Stacked
Autoencoders in Deep Learning. J. Med. Syst. 2020, 44, 32. [CrossRef] [PubMed]
33. Raja, P.S.; Rani, A.V. Brain tumor classification using a hybrid deep autoencoder with Bayesian fuzzy clustering-based segmenta-
tion approach. Biocybern. Biomed. Eng. 2020, 40, 440–453. [CrossRef]
34. Arunkumar, N.; Mohammed, M.A.; Ghani, M.K.A.; Ibrahim, D.A.; Abdulhay, E.; Ramirez-Gonzalez, G.; Albuquerque, V.H.C.
K-Means clustering and neural network for object detecting and identifying abnormality of brain tumor. Soft Comput. 2019, 23,
9083–9096. [CrossRef]
35. Arunkumar, N.; Mohammed, M.; Mostafa, S.A.; Ibrahim, D.A.; Rodrigues, J.J.; De Albuquerque, V.H.C. Fully automatic model-
based segmentation and classification approach for MRI brain tumor using artificial neural networks. Concurr. Comput. Pract.
Exp. 2020, 32, e4962. [CrossRef]
Diagnostics 2021, 11, 1589 19 of 19

36. Odusami, M.; Maskeliūnas, R.; Damaševičius, R.; Krilavičius, T. Analysis of Features of Alzheimer’s Disease: Detection of Early
Stage from Functional Brain Changes in Magnetic Resonance Images Using a Finetuned ResNet18 Network. Diagnostics 2021,
11, 1071. [CrossRef]
37. Huang, Z.; Xu, H.; Su, S.; Wang, T.; Luo, Y.; Zhao, X.; Liu, Y.; Song, G.; Zhao, Y. A computer-aided diagnosis system for brain
magnetic resonance imaging images using a novel differential feature neural network. Comput. Biol. Med. 2020, 121, 103818.
[CrossRef] [PubMed]
38. Rundo, L.; Militello, C.; Tangherloni, A.; Russo, G.; Vitabile, S.; Gilardi, M.C.; Mauri, G. NeXt for neuro-radiosurgery: A fully
automatic approach for necrosis extraction in brain tumor MRI using an unsupervised machine learning technique. Int. J. Imaging
Syst. Technol. 2018, 28, 21–37. [CrossRef]
39. Mekhmoukh, A.; Mokrani, K. Improved Fuzzy C-Means based Particle Swarm Optimization (PSO) initialization and outlier
rejection with level set methods for MR brain image segmentation. Comput. Methods Programs Biomed. 2015, 122, 266–281.
[CrossRef] [PubMed]
40. Han, C.; Rundo, L.; Murao, K.; Noguchi, T.; Shimahara, Y.; Milacski, Z.Á.; Koshino, S.; Sala, E.; Nakayama, H.; Satoh, S. MADGAN:
Unsupervised medical anomaly de-tection GAN using multiple adjacent brain MRI slice reconstruction. BMC Bioinform. 2021,
22, 1–20. [CrossRef]
41. Sahu, Y.K.; Pandey, C.; Biswas, P.; Khan, M.; Rathore, S. Minimum Time Delay and More Efficient Image Filtering Brain Tumour
Detection with the help of MATLAB. In Proceedings of the 2020 International Conference on Communication and Signal
Processing (ICCSP), Tamilnadu, India, 11–13 November 2020; pp. 1195–1199.
42. Ramesh, S.; Sasikala, S.; Paramanandham, N. Segmentation and classification of brain tumors using modified median noise filter
and deep learning approaches. Multimed. Tools Appl. 2021, 80, 11789–11813. [CrossRef]
43. Kleesiek, J.; Biller, A.; Urban, G.; Kothe, U.; Bendszus, M.; Hamprecht, F. Ilastik for multi-modal brain tumor segmentation. In
Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge); Spyridon: Boston, MA, USA, 2014; pp. 12–17.
44. Reza, S.M.S.; Mays, R.; Iftekharuddin, K.M. Multi-fractal detrended texture feature for brain tumor classification. In Medical
Imaging 2015: Computer-Aided Diagnosis; SPIE: Bellingham, WA, USA, 2015; Volume 9414, p. 941410.
45. Menze, B.H.; Jakab, A.; Bauer, S.; Kalpathy-Cramer, J.; Farahani, K.; Kirby, J.; Burren, Y.; Porz, N.; Slotboom, J.; Wiest, R.; et al. The
Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE Trans. Med. Imaging 2014, 34, 1993–2024. [CrossRef]
46. Badža, M.M.; Barjaktarović, M. Classification of Brain Tumors from MRI Images Using a Convolutional Neural Network. Appl.
Sci. 2020, 10, 1999. [CrossRef]
47. Toğaçar, M.; Ergen, B.; Cömert, Z. BrainMRNet: Brain tumor detection using magnetic resonance images with a novel convolu-
tional neural network model. Med. Hypotheses 2020, 134, 109531. [CrossRef] [PubMed]
48. Chavan, N.V.; Jadhav, B.; Patil, P. Detection and classification of brain tumors. Int. J. Comp. Appl. 2015, 112, 45–53.
49. Pereira, S.; Pinto, A.; Alves, V.; Silva, C. Brain Tumor Segmentation Using Convolutional Neural Networks in MRI Images. IEEE
Trans. Med. Imaging 2016, 35, 1240–1251. [CrossRef]
50. Hemanth, D.J.; Anitha, J.; Naaji, A.; Geman, O.; Popescu, D.E.; Son, L.H.; Hoang, L. A Modified Deep Convolutional Neural
Network for Abnormal Brain Image Classification. IEEE Access 2018, 7, 4275–4283. [CrossRef]
51. Ozyurt, F.; Sert, E.; Avcı, D. An expert system for brain tumor detection: Fuzzy C-means with super resolution and convolutional
neural network with extreme learning machine. Med. Hypotheses 2020, 134, 109433. [CrossRef]
52. Chelghoum, R.; Ikhlef, A.; Hameurlaine, A.; Jacquir, S. Transfer learning using convolutional neural network architectures for
brain tumor classification from MRI images. In Proceedings of the International Conference on Artificial Intelligence Applications
and Innovations, Crete, Greece, 25–27 June 2020; pp. 189–200.
53. Biratu, E.S.; Schwenker, F.; Debelee, T.G.; Kebede, S.R.; Negera, W.G.; Molla, H.T. Enhanced Region Growing for Brain Tumor MR
Image Segmentation. J. Imaging 2021, 7, 22. [CrossRef] [PubMed]
54. Saba, T.; Mohamed, A.S.; El-Affendi, M.; Amin, J.; Sharif, M. Brain tumor detection using a fusion of handcrafted and deep
learning features. Cogn. Syst. Res. 2020, 59, 221–230. [CrossRef]
55. Wu, W.; Chen, A.Y.C.; Zhao, L.; Corso, J.J. Brain tumor detection and segmentation in a CRF (conditional random fields)
framework with pixel-pairwise affinity and superpixel-level features. Int. J. Comput. Assist. Radiol. Surg. 2014, 9, 241–253.
[CrossRef]
56. Bauer, S.; Fejes, T.; Slotboom, J.; Wiest, R.; Nolte, L.-P.; Reyes, M. Segmentation of brain tumor images based on integrated
hierarchical classification and regularization. In Proceedings of the MICCAI BraTS Workshop; Miccai Society: Nice, France, 2012; p. 11.
57. Huang, M.; Yang, W.; Wu, Y.; Jiang, J.; Chen, W.; Feng, Q. Brain tumor segmentation based on local independent projection-based
classification. Trans. Biomed. Eng. J. 2014, 61, 2633–2645. [CrossRef] [PubMed]
58. Amin, J.; Sharif, M.; Yasmin, M.; Fernandes, S.L. Big data analysis for brain tumor detection: Deep convolutional neural networks.
Future Gener. Comput. Syst. 2018, 87, 290–297. [CrossRef]
59. Khan, M.A.; Ashraf, I.; Alhaisoni, M.; Damaševičius, R.; Scherer, R.; Rehman, A.; Bukhari, S.A.C. Multimodal Brain Tumor
Classification Using Deep Learning and Robust Feature Selection: A Machine Learning Application for Radiologists. Diagnostics
2020, 10, 565. [CrossRef] [PubMed]