International Journal of Neuroscience

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/ines20

Detection of brain lesion location in MRI images

using convolutional neural network and robust
PCA

Mohsen Ahmadi , Abbas Sharifi , Mahta Jafarian Fard & Nastaran Soleimani

To cite this article: Mohsen Ahmadi , Abbas Sharifi , Mahta Jafarian Fard & Nastaran Soleimani
(2021): Detection of brain lesion location in MRI images using convolutional neural network and
robust PCA, International Journal of Neuroscience, DOI: 10.1080/00207454.2021.1883602

To link to this article: https://doi.org/10.1080/00207454.2021.1883602

Published online: 12 Feb 2021.

Submit your article to this journal

Article views: 238

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=ines20
INTERNATIONAL JOURNAL OF NEUROSCIENCE
https://doi.org/10.1080/00207454.2021.1883602

ORIGINAL ARTICLE

Detection of brain lesion location in MRI images using convolutional neural

network and robust PCA
Mohsen Ahmadia , Abbas Sharifib , Mahta Jafarian Fardc and Nastaran Soleimanid
a
Department of Industrial Engineering, Urmia University of Technology, Urmia, Iran; bDepartment of Mechanical Engineering, Urmia
University of Technology, Urmia, Iran; cDepartment of Electrical Engineering, Islamic Azad University Science and Research, Razavi
Khorasan, Iran; dDepartment of Electronics and Telecommunications (DET), University of Politecnico di Torino, Turin, Italy

ABSTRACT ARTICLE HISTORY

Purpose and aim: Detection of brain tumors plays a critical role in the treatment of patients. Received 29 May 2020
Before any treatment, tumor segmentation is crucial to protect healthy tissues during treatment Revised 15 October 2020
and to destroy tumor cells. Tumor segmentation involves the detection, precise identification, Accepted 16 November 2020
and separation of tumor tissues. In this paper, we provide a deep learning method for the seg-
KEYWORDS
mentation of brain tumors. CNN; MRI; brain tumor;
Material and methods: In this article, we used a convolutional neural network (CNN) to seg- robust PCA; segmentation
ment tumors in seven types of brain disease consisting of Glioma, Meningioma, Alzheimer’s,
Alzheimer’s plus, Pick, Sarcoma, and Huntington. First, we used the feature-reduction-based
method robust principal component analysis to find tumor location and spot in a dataset of
Harvard Medical School. Then we present an architecture of the CNN method to detect
brain tumors.
Results: Results are depicted based on the probability of tumor location in magnetic resonance
images. Results show that the presented method provides high accuracy (96%), sensitivity
(99.9%), and dice index (91%) regarding other investigations.
Conclusion: The provided unsupervised method for tumor clustering and proposed supervised
architecture can be potential methods for medical uses.

1. Introduction treatment, tumor segmentation is critical to protect

healthy tissues during treatment and to destroy tumor
Cancer can be defined as the uncontrolled growth and
cells. Tumor segmentation involves the detection, pre-
division of body cells. This leads to abnormal cell devel-
opment and separation in the tissue of the brain as a cise identification, and separation of tumor tissues. This
tumor. This event is called a brain tumor. A brain tumor means that image segmentation involves manually
is an irregular brain injury that may be cancerous (malig- interpreting by a human expert and segmenting a large
nant) or non-cancerous (benign). Although brain tumors number of multi-state MRI images. Many kinds of
are not very common, they are one of the deadliest research have proposed computational methods for
types of cancers. Preliminary brain tumor detection plays tumor detection and segmentation since segmentation
a significant part in the likelihood of recovery and ther- is a time-consuming process and manual segmentation
apy. Techniques of medical imagery, such as computed has a possibility of error. High performance in segmen-
tomography (CT) imaging, positron emission tomography, tation using deep learning methods has made these
and single-photon emission CT, magnetic resonance methods a good option for achieving this goal.
spectroscopy, and magnetic resonance imaging (MRI) There are three approaches to image segmentation
offers useful information on the brain tumors’ form, including pixel sorting, tracking boundaries of change,
length, site, and metabolism that enables detect brain and the use of similarity and symmetry. The first
tumors as quickly as possible. Although these techniques method assumes that the pixels in each subclass have
are employed in collaboration to include the most accur- approximately fixed intensities which are usable for
ate brain tumor information, MRI is considered a well- structures with similar physiological properties. These
known imaging method due to its good contrast in soft algorithms may identify multiple components at the
tissue as well as its wide availability. Before any same time, but they are weak in noise and inadequacy

CONTACT Mohsen Ahmadi [email protected] Department of Industrial Engineering, Urmia University of Technology, Urmia, Iran.
ß 2021 Informa UK Limited, trading as Taylor & Francis Group
2 M. AHMADI ET AL.

in pictures. Conversely, methods that track the boun- connection, residual inception, and unpooling blocks
daries of change use intensity and spatial information. which leads a deeper and wider network.
Therefore, a subclass must be uniform and enclosed Milletari et al. [11] presented an approach based on
within a boundary of change. The third method is to Hough voting to execute segmentation by leveraging
use border similarity and symmetry using local histo- the abstraction capabilities of CNNs. Their architecture
grams. Diverse machine learning classification, seg- was used the CNN classification results and was car-
mentation, and clustering have been used in literature ried out voting by taking advantage of the features
during the last years. Due to the high ability of feature produced by the deepest portion of the network.
extraction of big dataset pictures and ‘no human Without the need for post-processing, the key to their
interference’, deep learning followed by the convolu- design was to use fewer training data and offer smooth
tional neural network (CNN) has got considerable segmentation contours. This method is pretty good for
attention in recent years. CNN based on deep learning big datasets. The exploration of small kernels in the
concepts have more convolutional and hidden layers segmentation images may also decrease the number of
and is more powerful in the segmentation of pictures weights in the network and improve the processing
time consequently. This task has been done by Pereira
in comparison to conventional neural networks.
et al. [12] they also used intensity normalization as a
In this article, we address a CNN architecture for
preprocessing step. These adjustments managed to
the segmentation of brain tumors for the seven brain
make MRI images effective for brain tumor segmenta-
diseases. First, the ground truth images are extracted
tion. Hassantabar et al. [25] presented CNN architec-
from original images using robust principle component
tures for classification and segmentation of X-Ray
analysis (PCA). The results of segmentation performance
images of patients lung involved in the 2019 corona-
analysis are presented in the following section.
virus. They used also a deep neural network with a
fractal feature method for feature extraction of the
2. Related works images. Their method outperformed other approaches
in the diagnosis and detection field (Table 1).
The deep learning method is one of the interesting
approaches between scholars. It has been employed
in images, voice recognition, and identification of 3. The method and materials
genotype/phenotype and classifying of diseases, such 3.1. Proposed method
as target detection and segmentation. CNN is one of
the well-known methods is classification and segmen- In this article, we detect brain tumor location using
deep learning methods. The first step is the extraction
tation. CNN-based segmentation has been applied in
of ground truth images from main images of seven
many fields with a wide variety of applications ranging
types of brain lesions. In this step, we presented a
from image retrieval to object recognition such as con-
new type of clustering method as Robust PCA for the
ditional random fields in pattern and image recognition
extraction of tumor spots from other parts of images.
[1], objective detection, segmentation and localization
After clustering images, we used enhancement meth-
[2], and medical applications such as CT images [3],
ods to remove little spots from images. Then, the
breast tumors classification and detection [4], X-Ray
main images are allocated as the input layer of the
image segmentation [5], Prostate segmentation from
CNN and ground truth images as the output layer.
MRI [6], and widely used in brain tumors detection and The training process is performed and consequently,
segmentation from divers analysis images [7–9]. Among results are recorded. The block diagram of the pre-
these utilities, CNN-based segmentation of MRI images sented process is illustrated in Figure 1.
to detect brain tumors has been investigated increas-
ingly by researches for over past decade.
Chen et al. [10] suggested CNN architecture to fix 3.2. Robust principal component analysis
some of the problems of image segmentation, such as (robust PCA)
indistinguishable features learned from standard con- One of the most important findings of linear algebra
volution layers, where variations between components is the analysis of principal components since the sim-
in terms of intensity, position, form, and size are sub- ple and non-parametric approach is to derive mean-
tle. They synthesized an architecture named Dense- ingful details from confusing sets. In the PCA, data are
Res-Inception Net to solve this problem. Their key depicted from a high-dimensional space to a low
modification was in three blocks named dense dimension, and in fact, the PCA is one of the methods
 INTERNATIONAL JOURNAL OF NEUROSCIENCE 3

Table 1. Recent works on the CNN- and GAN-based segmentation.

Author Year Key work Type Method Ref.
Havaei et al. 2016 Applying two pathway architecture based on both local Brain tumor CNN [13]
details and global context
Shakeri et al. 2016 Operation a CNN-based model on a full-blown 2D Sub-cortical structures CNN [14]
image without any arrangement at testing time of the brain
Pan et al. 2015 To exploit the learning power of the deep learning Brain tumor CNN [15]
machine, apply the grading output on the test data
measured by the sensitivity and specificity
Fritscher et al. 2016 Using a patch-based segmentation to improve Brain tumor CNN [3]
the accuracy
Nema et al. 2020 Using a RescueNet based on the residual and Brain tumor GAN [16]
mirroring principles.
Rezaei et al. 2019 Using a voxel-GAN approach to improve the issue of Brain tumor GAN [17]
unbalanced data for the tumor segmentation
Chen et al. 2019 Applying an end-to-end GAN-based model Brain tumor GAN [18]
Rundo et al. 2020 Using diverse architecture such as SegNet, U-Net, and Prostate cancer CNN [19]
pix2pix to segmentation the MRI images for
prostate cancer
Kamnitsas et al. 2017 Using an 11 layer 3D convolutional architecture and a Brain lesion CNN [20]
dual pathway for multi-scale processing
Isola et al. 2017 Changed the classic GAN as the generator and the Multi images GAN [21]
convolutional patch GAN classifier as the
discriminator with U-net-based architecture
Badrinarayanan et al. 2017 A modified encoder-decoder architecture called SegNet Multi images CNN [22]
has been suggested
Xue et al. 2018 An updated GAN called SegAN with a completely Medial images GAN [23]
convolutional neural network was suggested as the
segmentor to produce segmentation label maps

squares error, assuming that the vector of properties

is X 2 Rd (d-dimensional space).
In Robust PCA for segmentation, the main process
is abating of parameters L and S from the methods.
These parameters are defined as follow:


1 Y
L ¼ D ,X  S þ (1)
l l


k Y
S ¼ S ,X  L þ (2)
l l
where X is input m  m image, l is the augmented
Lagrangian parameter, k is the regularization param-
eter. The functions D is the shrinkage operator for sin-
gular values of the input image and S is shrinkage
operator that defined follow:
Dðs, X Þ ¼ uv  Sðs, sÞ (3)
X
Sðs, X Þ ¼ maxðj X j  s, 0Þ (4)
jXj

Figure 1. Conceptual diagram of the presented approach. where u and v and s are the singular parameters of a
symbolic matrix. The numeric unitary matrices consist
of selection of features and is required to decrease of u as singular vectors, v diagonal matrix, and s sin-
the dimension to explore the features in a simplified, gular values. Moreover, Y and Z are augmented
lower-dimensional space. What PCA does is that it Lagrangian multiplier is below:
obtains a linear transformation that takes the vector of
Z ¼X LS (5)
the h-dimensional properties to the d-dimensional X
properties of the vector (h < d), so that the informa- Y¼ lZ (6)
tion is kept almost completely (maximum possible
information) and the least-squares mean error The The iteration is repeated until E is reached to
PCA can be obtained by the minimization of the least- adequate value as follow:
4 M. AHMADI ET AL.

|Z|
E¼ < 105 (7)
|X|

3.3. CNN
In this section, we explain the CNN deep learning
approach. This type of neural network is one of the
learning networks inspired by the Perceptron neural
network. This deep network contains an input layer, an
output layer, and a hidden deep layer. First, images or
data of the problem are categorized into the algorithm
and trained. Then the hidden output layer weights can
appear in several ways [24]. If the output of the algo-
rithm contains several numerical elements such as bin- Figure 2. Plots of some types of activation functions.
ary number or index then the proposed algorithm is a
classification or detection algorithm. While, if the out- the device becomes unusable when using these
put layer is the matrix as large as the input image as functions in most situations (Figure 2) [25].
ground truth images, the presented method is segmen-  Tan h function
tation or detection. CNN consists of several hidden sub- The tan h function is similar to Sigmoid. The out-
layer forms that are described as follows: put interval of the tan h function is between 1
and 0 and, like Sigmoid, its curve is S-shaped. The
 Convolutional layer advantage of tan h to Sigmoid is that in tan h, the
The center of the convolution network is the con- negative input values are mapped to negative val-
volution layer, and its output can be represented ues, and the input value of zero is mapped to near
as a three-dimensional neuron pile. In simple terms, zero. However, in sigmoid, negative values are writ-
a three-dimensional pile is the output of this sheet. ten to values close to zero. The tan h function is
The CNN network employs separate kernels in also derivative and monotonically ascending.
these layers to convolve the input image as well as  ReLU
the core function maps. There are three major The rectified linear unit ReLU functionality is one of
advantages to the working of convolution [25]: the activation functions introduced in recent years.
 In each function map, the weight sharing pro- ReLU is a function of activation that extends to all
cess leads to a sharp decline in the number components. Its aim is to provide nonlinear behav-
of parameters. ior to the network. This functionality was imple-
 The local relation knows the correlation mented by Krizhevsky et al. in 2012. This function
between the pixels of the neighbor. reaches all pixels in the image and renders all
 A variation in the position of the object cre- negative values zero, as can be seen in Formula 1.
ates stability. The aim of using ReLU is to describe a nonlinear
 Activation functions component and its nonlinear training of the convo-
Commonly, activation mechanisms are imple- lution neural network (Convolution is a linear
mented in the neural network to achieve the method that is accomplished by the elements
desired response from the input functions. For being multiplied and summed up) [25].
neural networks, various activation functions may This function is the most common activation func-
be used; the sigmoid and hyperbolic tangent acti- tion used in the CNN. The advantages of this func-
vation functions are the most relevant of them. tion are:
Input that may have a value between þ1 and –1  The gradient is not saturated in the positive region
is obtained by the sigmoid function and the output  The calculations for the threshold requirements
is the interval between 0 and 1 [25]. An output are simple.
value between 1 and 1 is given by the hyperbolic  Works faster than sigmoidal and hyperbolic tangent
tangent function. This functionality makes these function and reduces the training error rate
two features on CNN networks less widely used. Disadvantages of this function are:
Because multiple values are used in the matrix  Outputs have no center of zero and values are
images, this outcome is the lack of image data and always positive
 INTERNATIONAL JOURNAL OF NEUROSCIENCE 5

 Gradients disappear for values smaller than zero, the network to classify features with more image
and new techniques such as RReLU, PReLU, and space. During the subtracting sampling task, pool-
LeakyReLU are used to overcome this problem. ing in the CNN is required to summarize the fea-
Leaky ReLU For positive inputs it is the same as for ture, so we can get into deeper network layers.
the previous function, but for inputs smaller than The spatial information storage space reduces as a
zero it acts as f(x) ¼ 0.01x. It reduces the negative result of the sampling as we approach the end of
numbers to very small (but effective) use. Its applica- each step and wish to reduce the sampling. So we
tion is similar to the previous function. It is oblivious want to begin pooling to gather what we have to
in the figure that the slope is shifted from x ¼ 0 to the retain this data. Max and average are the two most
left, creating a leak and expanding the ReLU range. common ways of pooling [25].
 SoftMax There is also much debate within the researches
One of the most used functions in the classification about how to improve each of the max and aver-
is the SoftMax function. In some cases, the SVM age techniques. The difference between the two is
function is also used. But since the SoftMax func- negligible, so one of the common paradigms is
tion gives us a more accurate result of the prob- using max-pooling across the network to retain the
ability of a class, it is, therefore, more suitable for best features and average pooling at the endpoints
classification. The probability of each class is calcu- to get the final representation vector of the fea-
lated from the following formula. tures before the last dense layer, and then deliver
efyi Cefyi efyi þ logC everything to SoftMax.
Pðyi xi ; W Þ ¼ P fj ¼ P fj ¼ P fj (8)  Padding
je C je j e þ logC
Sometimes the size of the output matrix needs to
be controlled. An extra layer (with zero value) is
Here yi is the correct class for xi image. Adding a
added to all around the matrix in order to get an
constant coefficient such as C to both sides does
not make a difference in the answer, but it makes output matrix as same as the size of the input
the analysis easier. In this layer, a common value for matrix. This job is called padding, same padding or
this coefficient is assumed to be logC ¼  maxj fj : ‘zero paddings’. After applying the zero-padding
 Dropout and the filtration, the output image should have
Deep neural networks with a lot of parameters are the same size as the original image.
powerful systems for learning. However, saturation  Data augmentation
occurs on some of these networks, which is the One of the most often overlooked issues is data
case on large networks that require a long time to preparation, preprocessing, and data enhancement.
train and test. Dropout is a technique that is cur- However, this task is not always necessary. Before
rently widely used and is usually added as an extra running any kind of data processing, you must first
layer after other layers. This layer works by ran- check whether your task needs to be preprocessed.
domly removing some units during network train- For example, in the classification of images, and
ing. To do this task, consider a probability value of implemented a standardized protocol for mean-nor-
P, and the matrix is randomly generated as the malization of images is based on the mean of the
input length. In this matrix, each value that has a trained data. Researches have repeatedly shown
probability more than the expected probability is that average normalization is the best thing to do
kept and the other values are removed from the for preprocessing. On the other hand, to optimize
matrix. Dropout is a technique that prevents the the images, mean-normalization can damage the
network from being ‘over-fit’. As its name implies, network and produce less accurate results. Any task
during the learning, some neurons are released by that is associated with very subtle differences in fea-
chance. This means that learning takes place on tures such as color, appearance, the overall shape,
different architectures with different sets of neu- and semantic differences in the image is likely to
rons. Dropout can be considered as a group tech- benefit from the lack of average normalization.
nique in which the outputs of multiple networks Data Augmentation, on the other hand, is closely
are blended to form the final output. related to uniform performance enhancement both
 Max pooling in absolute accuracy and in the generalization of
There are several effects of the use of max-pooling the network. This is done in all types of tasks, such
in neural networks. The use of max pooling enables as high-level classification to low-level optimization.
the network, with only small adjustments to the This means that the type of data augmentation
image, to first define the object. Secondly, it helps used must be under consideration.
6 M. AHMADI ET AL.

Figure 3. The results of the detection of tumor location: (a) original image, (b) the result of Robust PCA, (c) the result of image
enhancement.

4. Results images. This method used single-channel images for

clustering the pixels of the image to two separate
4.1. Database
classes. With the use of Robust PCA, we reached to
We used seven brain diseases in this article to intro- ground truth images. The results of this clustering
duce and validate the approaches proposed. They algorithm are shown in Figure 3(b).
comprise of Glioma, Huntington, Meningioma, Select, Figure 3 shows the results of determining ground
and Sarcoma, Alzheimer’s, Alzheimer’s with visual truth images. To enhance the exact and clear loca-
agnosia. 80 MRI images from the Harvard Medical tion of the tumor, we extract only blobs larger than
School website [26] are included in the image. In the 150 pixels and removed the white small blobs. The
axial plane, all images come from T2-weighted MR resulting image is an acceptable ground truth image
brain images and have 256  256 pixels. The photos (Figure 3(c)). This two-step method determined the
are in a three-channel GIF format. segmented images very quickly and was used for
the seven types of brain disease with different l
4.2. Preprocessing and k values. For training images using CNN, we
need a huge amount of data to reached accurate
Before starting the CNN segmentation process, some results. Moreover, preventing from overfitting we
preprocesses are required. The input data is packages need data augmentation. For the increasing number
of images with GIF format; therefore, the first step is of images transform original images. We used the
converting the dataset to RGB and then grayscale Gaussian function for adding noises, histogram
images. These steps transform three-channel images equalization to enhance image contrasts, median fil-
into single-channel images. The input images are mix- ter, rotation in 90 degrees. We used these methods
tures of seven types of brain disease; therefore, the to produce input images. After augmentation of the
presented CNN method can be adopted to seven dif- original image, we have 1120 input and 1120
ferent types of brain MRI images. Second, preprocess- ground truth images.
ing is the achievement of ground truth images. The
ground truth images are the matrixes that determine
the tumor location in binary format. Such that, the
4.3. Implementation of CNN architecture
elements of matrixes in the location of tumor results 1 To begin with, we present a deep architecture for a
and other points equal 0. The ground truth images CNN. Three convolutional layers are used in this archi-
are determined in different ways, such as manual tecture. The diagram of the CNN layers is presented in
selection by physicians or using some images process- Figure 4.
ing mathematical algorithms. In this article, 1120 images were used for training
In this article, we used the Robust PCA method to the CNN architecture, and 80 images were used for
reach the best ground truth images. Although the testing the resulted weights. The first layer is a convo-
Robust PCA typically is used for feature reduction or lutional layer with the use of 32 filters in size of 3  3
determining outliers and special leverages, we used with stride 1 in each direction. Also, the ReLU activa-
this method for detection of brain location in MRI tion function is used for removing negative results.
 INTERNATIONAL JOURNAL OF NEUROSCIENCE 7

Figure 4. The architecture of presented CNN layers.

Figure 5. (a) Original image, (b) ground truth image, (c) CNN segmentation result.

Then use a dropout of 50%for dropping some items. detection is higher than other pixels. Therefore, we
The next layer is the Maxpooling function in 2  2 calculate the probability of detection in each pixel.
matrixes with stride 2 in each direction. The second The resulted images for seven brain diseases are pre-
convolutional layer consists of 64 (3  3) filters and a sented in Figure 6. The left contour shows the prob-
transposed convolutional layer is brought with 32 ability of tumor on images. So, the prediction presents
(4  4) filters. The final convolutional layer used two reliable results.
(1  1) filters to increase accuracy. The final step in the
segmentation layers is SoftMax. Segmentation of
images almost is the classification of pixels with 5. Discussion
ground truth images labels. The training process is
5.1. Performance analysis
performed with a training rate of 0.01 with 500
epochs. The resulted images are shown in Figure 5. In order to determine the outcome of binary classifi-
The process is done for seven different brain diseases. cation (duality) of pixels, sensitivity, and specificity
Because of the better description of the results in in the statistics of the two metrics are considered.
regions with accumulated pixels, the probability of The consistency of the outcomes of a test that
8 M. AHMADI ET AL.

Figure 6. Results of segmentation using CNN, left: original images, middle: ground truth by robust PCA, right: CNN results.

separates the information into these two types is  True positive (TP): The white pixel has been accur-
observable and informative using sensitivity and ately detected.
attribute metrics where the results can be separated  False positive (FP): The black pixel is
into positive and negative classes. Sensitivity means detected inaccuracy.
the number of positive cases that would be accur-  True negative (TN): The black pixel is
ately checked as positive. Specificity means the detected accurately.
number of negative cases that accurately label them  False negative (FN): The white pixel is
as negative [27]. detected inaccuracy
 INTERNATIONAL JOURNAL OF NEUROSCIENCE 9

Figure 6. Continued

Figure 7. Accuracy of the test dataset.

The sensitivity of separating the percentages of TP

cases into the sum of TP and FN cases in mathemat-
ical terms. Figure 8. Pairwise images of GT and CNN.

TP employed. Nevertheless, it is not possible to interpret

Sensitivity, TPR ¼ (9)
TP þ FN the outcome of a test utilizing sensitivity and specifi-
Likewise, the specificity leads in the classification city alone.
into the number of FP and TN cases of TN cases. TP
Precision, PPV ¼ (11)
TP þ FP
TN
Specificity, TNR ¼ (10) TP þ TN
TN þ FP Accuracy ðACCÞ ¼ (12)
TP þ TN þ FP þ FN
The sensitivity of the test and its specificity depend FP
Fall  out ðFPRÞ ¼ (13)
on the quality of the test and the type of test FP þ FN
10 M. AHMADI ET AL.

Table 2. The average of performance criteria values for the

CNN method.
Miss
Criteria AUC Fallout rate Dice Sensitivity Specificity Accuracy
Mean 0.70 0.002 0.001 0.912 0.999 0.998 0.96
Value

Table 3. The comparison of the presented architecture with

other methods.
Reference Method of Detection Ground truth Dice index
[28] Experimental Manual 0.88
[29] SVM Semi-automatic 0.86
[30] CNN K-means 0.83
[31] GAN Automatic 0.8
[16] Unpaired GAN Automatic 0.94
[12] CNN Automatic 0.78
[32] 2D cGAN Automatic 0.59
[32] 32 cGAN Automatic 0.72
Presented CNN Robust PCA 0.91
Figure 9. ROC curve for the presented approach. method

points that exist in the GT but not predicted properly.

In other words, green pixels show the TP and white
pixels show the FN pixels. Also, Purple points are the
paints that are predicted as a tumor spot incorrectly,
which is names FP. Other back points show the TN.
Figure 9 shows the ROC curve of images used in this
article for segmentation using the CNN method. This
chart shows the TP rate versus the FP rate. Regarding
results, if the curve inclines to TP rate ¼ 1 and FP rate
¼ 0, the presented method reveals high performance.
Each of the curves is belonging to an image. Usually,
the ROC curve for better description is illustrated as
under the curve area or AUC. For every image with a
high AUC value, the presented method results in high
Figure 10. The area under the ROC curve (AUC) for the pre- performance. The AUC values are depicted on the
sented approach. chart in Figure 10. Based on the findings of Figure 10,
the AUC value for all of the images is averaged to
FN 0.7031 which is an acceptable value for the
Missrate ¼ (14)
TP þ FN CNN method.
TN Total performance analysis values are summarized
J¼ (15)
TN þ FP þ FN as following Table 2. Regarding these findings total
2J accuracy is 96% that shows the high performance of
Dice ¼ (16)
1 þ J this method. The values of specificity, sensitivity, and
The ACC of the test images are shown in Figure 7. dice index are 99.8%, 99.9%, and 91.2%, respectively.
ACC is tolerated from 92% to 99%. The higher ACC Table 3 shows the comparison of the presented
corresponds to 0–30 or Alzheimer’s, Alzheimer’s plus, method using the Dice index. Results show that the
and Glioma almost 97%. Moreover, ACC of 60–80 has presented method shows a high dice index and
high accuracy. These diseases are detected with high acceptable architecture.
accuracy. However, the ACC of 30–60 is corresponded
to Huntington, Meningioma, and pick. The pairwise
6. Conclusion
image of one case is depicted in Figure 8.
In Figure 8, one of the images of Sarcoma is pre- Automated detection of brain disease plays a critical
sented. In this figure, the green spots show the inter- role in the treatment of patients. So, investigations in
section of GT and CNN and white color pixels are the the fields of medical image processing are the next
 INTERNATIONAL JOURNAL OF NEUROSCIENCE 11

step in medical technology. In this article, we pre- [2] Cai Z, Vasconcelos N. Cascade R-CNN: High-Quality
sented two approaches to the detection of seven Object Detection and Instance Segmentation. arXiv
preprint arXiv:1906.09756, 2019.
types of brain lesions. The first method is the Robust
[3] Fritscher K, Raudaschl P, Zaffino P, et al. Deep neural
PCA. This method was used as clustering data and fea- networks for fast segmentation of 3D medical images.
ture reduction in recent researches. We used this In: International Conference on Medical Image
property of robust PCA to separate lesion area from Computing and Computer-Assisted Intervention,
brain MRI images. The Robust PCA cluster input 2016. p. 158–165.
[4] Rouhi R, Jafari M, Kasaei S, et al. Benign and malig-
images to two separate black and white clusters. This nant breast tumors classification based on region
clustering process almost can detect tumor location. growing and CNN segmentation. Expert Syst Appl.
Therefore, the other small hit points are enhanced 2015;42(3):990–1002.
with the use of algorithms that remove small spots. [5] Bullock J, Cuesta-Lazaro C, and Quera-Bofarull A.
XNet: a convolutional neural network (CNN) imple-
The resulting images are used as ground truth images
mentation for medical x-ray image segmentation suit-
to the next step. The second methods are a CNN. We able for small datasets. In: Medical Imaging 2019:
used result images of robust PCA in the CNN algo- Biomedical Applications in Molecular, Structural, and
rithm. The input layer of the CNN method is a folder Functional Imaging, 2019. p. 109531Z.
of gray MRI images with a size of 256  256. On the [6] Zhu Q, Du B, Turkbey B, et al. Deeply-supervised CNN
for prostate segmentation. In: 2017 International Joint
other hand, the output layer is consisting of the Conference on Neural Networks (IJCNN), 2017, p.
ground truth images. For the increasing number of 178–184.
images and preventing from overfitting, we used the [7] Milletari F, Navab N, and Ahmadi S-A. V-net: Fully
data augmentation approach. With the use of rotation, convolutional neural networks for volumetric medical
image segmentation. In: 2016 Fourth International
noise, and filtration the numbers of images are
Conference on 3D Vision (3DV), 2016. p. 565–571.
increased to 2240 images. Results are depicted based [8] Kayalibay B, Jensen G, Smagt PVD. CNN-based seg-
on the probability of tumor location in MRI images. mentation of medical imaging data. arXiv preprint
The higher ACC corresponds to 0-30 or Alzheimer’s, arXiv:1701.03056, 2017.
Alzheimer’s plus, and Glioma almost 97%. Moreover, [9] Moeskops P, Wolterink JM, van der Velden BHM, et al.
Deep learning for multi-task medical image segmen-
ACC of 60–80 has high accuracy. These diseases are tation in multiple modalities. In: International
detected with high accuracy. However, the ACC of Conference on Medical Image Computing and
30–60 is corresponded to Huntington, Meningioma, Computer-Assisted Intervention, 2016. p. 478–486.
and pick. Results show that the presented method [10] Chen L, Bentley P, Mori K, et al. Drinet for medical
image segmentation. IEEE Trans Med Imaging. 2018;
provides higher accuracy (96%), sensitivity (99.9%),
37(11):2453–2462.
and dice index (91%) than other researches. [11] Milletari F, Ahmadi S-A, Kroll C, et al. Hough-CNN:
deep learning for segmentation of deep brain regions
in MRI and ultrasound. Comput Vis Image Underst.
Disclosure statement 2017;164:92–102.
No potential conflict of interest was reported by the author(s). [12] Pereira S, Pinto A, Alves V, et al. Brain tumor segmenta-
tion using convolutional neural networks in MRI images.
IEEE Trans Med Imaging. 2016;35(5):1240–1251.
Funding [13] Havaei M, Davy A, Warde-Farley D, et al. Brain tumor
segmentation with deep neural networks. Med Image
The funding sources had no involvement in the study Anal. 2017;35:18–31.
design, collection, analysis or interpretation of data, writing [14] Shakeri M, Tsogkas S, Ferrante E, et al. Sub-cortical
of the manuscript, or in the decision to submit the manu- brain structure segmentation using F-CNN’s. In 2016
script for publication. IEEE 13th International Symposium on Biomedical
Imaging (ISBI), 2016. p. 269–272.
[15] Pan Y, Huang W, Lin Z, et al. Brain tumor grading
ORCID based on neural networks and convolutional neural
Mohsen Ahmadi http://orcid.org/0000-0003-1550-110X networks. In 2015 37th Annual International
Abbas Sharifi http://orcid.org/0000-0002-2915-2740 Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC), 2015. p. 699–702.
[16] Nema S, Dudhane A, Murala S, et al. RescueNet: an
References unpaired GAN for brain tumor segmentation. Biomed
Signal Process Control. 2020;55:101641.
[1] Liu F, Lin G, Shen C. CRF learning with CNN features [17] Rezaei M, Yang H, and Meinel C. voxel-GAN:
for image segmentation. Pattern Recognit. 2015; Adversarial Framework for Learning Imbalanced Brain
48(10):2983–2992. Tumor Segmentation: 4th International Workshop,
12 M. AHMADI ET AL.

BrainLes 2018, Held in Conjunction voxel-GAN: [27] Hamzenejad A, Jafarzadeh Ghoushchi S, Baradaran V,
Adversarial Framework for Learning Imbalanced Brain et al. A robust algorithm for classification and diagno-
Tumor Segmentation, 2019. sis of brain disease using local linear approximation
[18] Chen H. Brain tumor segmentation with generative and generalized autoregressive conditional heterosce-
adversarial nets. 2019 2nd International Conference dasticity model. Mathematics. 2020;8(8):1268.
on Artificial Intelligence and Big Data (ICAIBD), 2019. [28] Menze BH, Jakab A, Bauer S, et al. The multimodal
p. 301–305. brain tumor image segmentation benchmark (BRATS).
[19] Rundo L, Han C, Zhang J, et al. CNN-based prostate IEEE Trans Med Imaging. 2015;34(10):1993–2024.
zonal segmentation on T2-weighted MR images: a [29] Havaei M, Larochelle H, Poulin P, et al. Within-brain
cross-dataset study. In Neural approaches to dynamics classification for brain tumor segmentation. Int J
of signal exchanges. Springer; 2020. p. 269–280. Comput Assist Radiol Surg. 2016;11(5):777–788.
[20] Kamnitsas K, Ledig C, Newcombe VF, et al. Efficient [30] Dvorak P, Menze B. Structured prediction with convo-
multi-scale 3D CNN with fully connected CRF for lutional neural networks for multimodal brain tumor
accurate brain lesion segmentation. Med Image Anal. segmentation. In: Proceedings of the Multimodal
2017;36:61–78. Brain Tumor Image Segmentation Challenge; 2015. p.
[21] Isola P, Zhu J-Y, Zhou T, et al. Image-to-image transla- 13–24.
tion with conditional adversarial networks. In [31] Rezaei M, Harmuth K, Gierke W, et al. A conditional
Proceedings of the IEEE Conference on Computer adversarial network for semantic segmentation of
Vision and Pattern Recognition, 2017. p. 1125–1134. brain tumor. In International MICCAI Brainlesion
[22] Badrinarayanan V, Kendall A, Cipolla R. Segnet: A Workshop; 2017. p. 241–252.
deep convolutional encoder-decoder architecture for [32] Yu B, Zhou L, Wang L, et al. 3D cGAN based cross-
image segmentation. IEEE Trans Pattern Anal Mach modality MR image synthesis for brain tumor seg-
Intell. 2017;39(12):2481–2495. mentation. In 2018 IEEE 15th International
[23] Xue Y, Xu T, Zhang H, et al. Segan: Adversarial net- Symposium on Biomedical Imaging (ISBI); 2018. p.
work with multi-scale L1 loss for medical image seg- 626–630.
mentation. Neuroinformatics. 2018;16(3-4):383–392. [33] Sun Y, Zhou C, Fu Y, et al. Parasitic GAN for semi-
[24] Krizhevsky A, Sutskever I, Hinton GE. Imagenet classifi- supervised brain tumor segmentation. In 2019 IEEE
cation with deep convolutional neural networks. In International Conference on Image Processing (ICIP),
Advances in Neural Information Processing Systems. 2019. p. 1535–1539.
2012. p. 1097–1105. [34] Dong X, Lei Y, Wang T, et al. Automatic multiorgan
[25] Hassantabar S, Ahmadi M, Sharifi A. Diagnosis and segmentation in thorax CT images using U-net-GAN .
detection of infected tissue of COVID-19 patients Med Phys. 2019;46(5):2157–2168.
based on lung x-ray image using convolutional neural [35] Han C, Rundo L, Araki R, et al. Infinite Brain Tumor
network approaches. Chaos Solitons Fractals. 2020; Images: Can GAN-based Data Augmentation Improve
140(140):110170. Tumor Detection on MR Images? In Proceeding
[26] The Whole Brain Atlas. http://www.med.harvard.edu/ Meeting on Image Recognition and Understanding
AANLIB/. (MIRU), 2018.