NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.

NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

A Modified Resnet (MRestNet) based

Classification for Brain Tumor Dataset
Mr.R.Vinayagamoorthy
Ph.D, Research scholar
Reg. No: 17224012161117
vinayagamoorthy3@gmail

Dr.R.Balasubramanian
Professor
Department of Computer Science and Engineering
ManonmaniamSundaranarUniversity,Tirunelveli-12
[email protected]

Abstract
The brain is considered one of the most important organs in human. A tumor in the brain can lead to loss
of functionality. Approximately 18,020 deaths are occurred in 2020 due to brain tumors. It is possible to
minimize these cases if a brain tumor is detected earlier. Brain tumors are the most widely recognized
and forceful infection, prompting a short future and much aggravation. Patient’s survival rates can be 7750
improved by timely and accurate diagnosis. The principle technique for distinguishing brain tumor is to
investigate an MRI picture that gives definite data about brain shape and abnormality identification in
brain tumor tissue with the quick advancement of Deep Learning (DL), particularly the improvement of
PC vision innovation, programmed brain tumor recognition is proposed by utilizing ResNet. As the
RESNET network deepens, its ability to extract network characteristics improves through layer-to-layer
feature fusion. ResNet has proven to operate better than other deep learning models, as the research
team tested it in an acceptable timeframe, and compared several deep learning models, which resulted
in a higher accuracy with ResNet. In this paper, we propose another model called Modified Residual
Network (MResNet) to classify brain tumor image. The proposed model has been assessed utilizing
activation layer on an openly accessible brain tumor image dataset having selected feature
images.Taking a more optimistic view is also possible with regards to skipping training in some residual
block layers.There is no general consensus on the optimal number of layers required for neural networks,
since the number of layers required may depend on the complexity of the dataset. The network will not
be trained on layers that are unimportant and do not contribute to overall accuracy by including skip
connections instead of treating the number of layers as an important hyperparameter to tune. We can
use skip connections to make our neural networks dynamic to tune the number of layers during training
optimally. Our model requires extremely less computational power and has much better precision results
when contrasted with other pre-trained models. The proposed techniques performed well when
compared to existing techniques like LeNet, AlexNet, VGG Net 16, ResNet 18.
Key words: Deep CNN, LeNet, AlexNet, VGG Net 16, ResNet-18, IResNet.
eISSN1303-5150 www.neuroquantology.com
NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

DOI Number: 10.14704/NQ.2022.20.15.NQ88767 NeuroQuantology2022;20(15): 7750-7767

Introduction tumors to form in the skull, which can increase

Movement of the human limbs is pressure on the brain and adversely affect
controlled by the brain. The brain also health. Medical imaging plays an important role
determines a number of advanced mental in the early detection and classification of brain
activities, including memory, language, learning tumors, which in turn assists in selecting the
and intelligence. In addition to affecting a most convenient treatment method for
person’s health, brain tumors can also directly patients. There are several ways to categorize
damage the brain’s structure. It is possible for brain tumors.

Input (Brain Tumor Dataset)

Testing Images Training Images
Glioma Tumor 100 Images Glioma Tumor 826 Images
Meningioma Tumor 115 Images Meningioma Tumor 822 Images
Pituitary Tumor 74 Images Pituitary Tumor 827 Images
No Tumor 105 Images No Tumor 395 Images

Feature Extraction

Feature Selection

7751
Classification

Fig 1-Classification Process

Pituitary tumors are benign tumors that image into its squashed form. The method
form in the pituitary gland near the brain and Contrast, Homogeneity, Energy, Entropy, Sum
can affect hormone levels in the body. of Averages, Sum of Variance, Autocorrelation,
The most common type of brain tumor Standard Deviation, etc. GLCM, Shape
is the glioma, which grows within the brain Descriptors, and Histogram-Based Metrics are
itself as well as mixing with normal brain tissue. significant Feature Extraction Techniques that
This is called an intra-axial brain tumor because can mitigate dataset dimensionality.
the tumor grows within the substance of the Feature Selection
brain. A hybrid algorithm combines a Genetic
Located just inside the skull, Algorithm and a Boruta Algorithm. Genboruta is
meningioma is tumors that form on membranes a hybrid algorithm that uses the advantages of
that cover the brain and spinal cord. existing algorithms to meet the needs of the
Feature Extraction selection of optimal feature subsets from a
Extracting the significant features of an limited number of features. The feature
image allows us to more easily comprehend it. selection method is applied to the data
This process is known as feature extraction. extracted from feature extraction in order to
Feature extraction involves transforming an select the most appropriate features.
eISSN1303-5150 www.neuroquantology.com
NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

Specifically, the Genboruta algorithm is used to in document recognition applications in the

select the most significant features. mid-20th century, as Lecun et al. created it in a
Classification form that was close to its current form. CNNs
In the area of Deep Learning and then continued to succeed in the deep learning
computer vision, Image Classification is a crucial field following its success. By increasing the
and extensive research domain. An image is training samples, CNNs are able to achieve the
classified into one of a number of predefined desired results of feature learning and infinite
classes based on the labels assigned to it. It accuracy, resulting in a more precise and robust
includes image acquisition, image processing, model than conventional machine learning.CNN
image object detection, image segmentation, architecture uses convolutional filters to extract
feature extraction, and image object features, and as we go deeper, we are able to
classification. Radiologists gain increased extract more complicated features (spatial and
diagnostic efficiency and provide an improved structural information). Filters are
path for surgical treatment by accurately convolutioned with patterns to yield unique
classifying images, especially in the field of features, which are then used to train a
medical imaging and biomedical imaging [1]. network for classification [4].
The growing availability of labeled LeNet
training data, the advancement of powerful In [5] a convolutional classifier, a small
graphics processing GPU, the development of segment of a pixel is used as the input. A kernel
new methods to learn and solve complicated size of 5 * 5 is used to identify low-level 7752
applications during the past few years have features (e.g., edges).The downsampled data is
contributed to the success of CNNs in medical obtained by using a max-pooling layer with
image applications [2]. stride 2.Another convolutional layer and a max-
CNNs can learn the most useful features pooling layer are deployed to determine the
automatically; they have demonstrated notable higher-level image structures.Two fully
performance in computer vision, particularly in connected layers make up the network's last
segmentation and classification. Kernel, section: each layer has one neuron as its
pooling, fully connected, and softmax functions output, while the other has two neurons.The
are the convolutional layers that each input rectified linear unit has been used as a transfer
passes.There are many layers in CNN that use function on all the convolutional layers.Deep
convolution filters to transform their input to learning uses LeNet, one of the most popular
some extent.Convolutional networks consist of and widely used network architectures[6].In [8]
three main layers: convolutional (a set of this CNN, images are classified according to the
learnable filters), pooling (used to reduce type of tumor based on the LeNet.By modifying
overfitting and reduce the size of images), and the arrangement and number of layers of
fully connected (to combine spatial and channel LeNet, the accuracy of the system has been
features together) [3]. improved.
Literature Review AlexNet
Deep CNN The Alex Net CNN model is used to
CNNs are convolutional neural classify tumor types based on their high-level
networks used in visual image analysis and are rich features.An Alex Net-based model is used
supposed to require little preparation. They are to classify tumors by tuning.Convolutional
a kind of deep learning.Data can be arranged in layers are composed of five, and fully
a variety of arrays using this technique. It is connected layers consist of three.Dropout can
based on the biological functions of the human be used to solve the problem of overfitting.Alex
brain."LeNet" was the first deep network used et.al developed a pre-trained model of the Alex
eISSN1303-5150 www.neuroquantology.com
NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

Net network.In transfer learning, the Alex Net parameters.It’s used for learn hidden features
Convolutional Neural Network is one of the [14].
frameworks used to differentiate benign tumors ResNet
from malignant.In order to classify an image, it This study uses the ResNet -18 model,
must be labeled [9].The brain MRI image is which is composed of 3 * 3 filters with stride
analyzed after detecting a tumor for and pad sets of 1, and an average pooling layer
malignancy.For training and classification of with a 1 * 1 filter and a fully connected layer
brain MRI images, the CNN-based AlexNet with a softmax layer [17].ResNet [18] is ranked
transfer learning architecture is utilized [10].In among the most efficient network architectures
this architecture, we can see five convolutional for image recognition tasks, since it effectively
layers and three fully connected layers.[11] deals with the problem of vanishing gradients in
describes a study that used AlexNet to classify deep networks.Introducing residual-linking, a
and segment abnormalities.These methods aim form of identity mapping that enables "skipped
to determine if CNN is applicable in classifying connections" to be added to the outputs of the
subjects into normal or abnormal categories in stacked layers.Identity connections are not
the simplest form.In a similar way to AlexNet complicated and improve training processes
with ReLU and Softmax activations, Dropout is without adding any complexity to the
used for overfitting. network.The ResNet18 architecture [19]
VGG Net prevents gradient vanishing by using residual
A VGG-16 network is an architecture connections between layers.
consisting of 16 layers of convolutional neural Improved ResNet 7753
networks. It is considered to be one of the most The ResNet18 [16] architecture has 18
brilliant vision models available today. Rather layers, including 13 convolutional layers and
than having more parameters, the VGG-16 three fully connected layers. Each CL uses a 3*3
model architecture focuses on a convolution kernel.Shortcuts are used in ResNets in which
layer with a size of 3x3. The significance of this one or more layers are skipped, as a result of
version is its weights are freely to be hadat which a lower training loss is achieved whereas
theinternet and may be downloaded to apply in stacked layers are added to their outputs;
personal models and application. It is thought therefore, complex computing is not
for its simplicity in comparison to required.Two layers are omitted from the skip
differentmodel architectures [12].DL techniques connections of ResNet18.ResNet-18 [20] is fine-
based on the compound of pre-trained VGG-16 tuned based on our brain MRI dataset, and
CNNs to classify 3 kinds of brain tumors (i.e., three RNNs are substituted for the fourth layer
pituitary tumor,glioma, and meningioma). The of the network to improve classification
aim of this research is the utilization of gray performance.The fully connected layers are
level of co-occurrence matrix GLCM feature changed and the output reshaped based on the
images furthermore, the original images as number of output classes in the dilemma.By
contributions to CNNs. Two gray level of co- unfreezing all layers and updating all
occurrence matrix features images utilized parameters of the network, we start by training
(energy, contrast, etc) [13].The VGG 16 a new ResNet18 from scratch by unfreezing all
techniques image input size is 224 x 224 x 3. IT layers.Furthermore, the final dense layer is
contains 16 convolution layers with a fixed max reshaped to have the same number of inputs
pooling layers of 2 x 2 sizes and filters 3 x 3 sizes and outputs as the dataset's classes [21].
all through the network.VGG-16 technique is a Methodology
bignetwork, with around 138 million Brain tumors are among the most
dangerous of cancerous diseases. Research
eISSN1303-5150 www.neuroquantology.com
NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

shows that the number of cases of brain tumor selected using Genboruta. It is proposed to use
is increasing worldwide. The brain is a vital deep CNNs based on MResNet to automatically
organ that can directly lead to death. It is classify brain tumor MR images into three
therefore imperative to detect and treat these tumors.Transfer learning was used to develop a
diseases early.GLCM is used to extract texture CNN model to detect brain tumors
features, and connected regions to extract automatically using MRI images
shape features. Shape features are found Deep convolutional neural networks are
directly in the segmented images. For the an effective tool for classifying images. Among
calculation of texture and shape features, it is the most commonly used deep network is
the simplest and least complex method. A Resnet.A new enhanced deep learning
histogram of oriented gradient (HOG) is a framework is presented here for the
mathematical expression that estimates how classification of multimodal brain tumors.From
big and how oriented a gradient is at each pixel the feature layers, the features were
of an image. The feature selection method is extracted.Optimizing the extracted features was
applied to the feature extraction data in performed using the GenBoruta.Each network's
optimized method to determine the most selected features are aggregated, analyzed and
relevant features. Important features are classified using MResnet.

7754

Fig 2- Proposed Classification Architecture

Feature Extraction Shape Descriptors
This section describes how to decrease A shape descriptor is a mathematical
a dataset's dimensionality using Feature function used to describe the shape of an
Extraction Techniques, including Shape image. A descriptor is generally a set of
Descriptors, Histogram-Based Metrics, and numerical values used to describe a particular
Grey-level Co-occurrence Matrix (GLCM)[15]. shape. Descriptors are usually represented by
vectors. They can be used to perform further
eISSN1303-5150 www.neuroquantology.com
 NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

tasks once they have been extracted. For measured, or the vector of shape descriptors
instance, the distance between the vectors of can be passed to a classifier to determine which
shape descriptors in a matching task can be class the given shape belongs to.
Shape Descriptors
1. Mesh Volume

2. Voxel Volume

3. Surface Area

4. Surface Area to Volume

Ratio

5. Sphericity

6. Compactness 1

7. Compactness 2

8. Spherical Disproportion

7755
9. Major Axis Length

10. Minor Axis Length

11. Least Axis Length

12. Elongation

eISSN1303-5150 www.neuroquantology.com
 NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

13. Flatness

Table 1 Shape Descriptors features and Formula

Histogram -Based Metrics rectangles raised over discrete intervals, with
An illustration of the frequency the area of each rectangle equal to the
distribution of data is called a histogram. It can frequency of the observations within the
be used to represent the distribution of data in interval.
a set of measurements. Typical histograms of
digital images are formed from tabular
frequencies represented as neighboring
Histogram-Based Metrics
1 Energy

2. Total Energy

3 Entropy

4 Max Intensity
5 Min Intensity
7756

6 Mean Value

7 Mean Absolute Deviation

8 Robust Mean Absolute

Deviation (RMAD)

eISSN1303-5150 www.neuroquantology.com
 NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

9 Range

10 Root mean square (RMS)

11 Standard deviation

12 Uniformity

13 Variance

14 Skewness 7757

15 Kurtosis

Table 2 Histogram Based Metrics Features and Formula

Grey-Level Co-occurrence Matrix parameters have been chosen for GLCM
In the GLCM function, pixels construction in order to extract the most
with specific values and spatial relationships are significant textural features with a minimum
compared to determine how often they occur in number of correlated features.
an image. Through rigorous analysis, the

Grey-level co-occurrence matrix GLCM

1. Autocorrelation

eISSN1303-5150 www.neuroquantology.com
 NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

2. Joint Average

3. Cluster Prominence

4. Cluster shade

5. Cluster Tendency

6. Contrast
7758

7. Correlation

8. Difference Average

9. Difference entropy

10. Difference variance

11. Dissimilarity

eISSN1303-5150 www.neuroquantology.com
 NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

12. Joint Energy

13. Joint Entropy

14. Homogeneity

15. Information
measure of
correlation 1

16. Information
measure of
correlation 2
7759
17. Inverse difference
Moment

18. Maximum
probability
coefficient

19. Inverse Difference

Moment
Normalized (IDMN)

20. Inverse Difference

(ID)

21. Inverse Difference

Normalized (IDN)

eISSN1303-5150 www.neuroquantology.com
 NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

22. Inverse Variance

23. Maximum
Probability
24. Sum Average

25. Sum variance

26. Sum entropy

27. Sum of squares

7760
Table 3 GLCM Features and Formula

Features Optimization Using Gen Boruta grouping tests before selecting the small
This research relies on using a hybrid number of important variables. The use of a
model that combines both a Genetic Algorithm hybrid technique overcomes the disadvantages
and a Boruta Algorithm, with an original aim of of each individual one [15].
S.No Genboruta (Features)
1. Surface Area 2. Range

3. Flatness 4. Cluster shade

5. Skewness 6. Dissimilarity

7. Uniformity 8. Maximum Probability

9. Contrast 10. Variance

11. Correlation 12. Max Intensity

Table4 Selected Features

Modified ResNet and 3 fully connected layers.All convolutional
There are 18 layers in the ResNet18 layers use 3*3 kernels.The MResNet algorithm
architecture, including 13 convolution layers makes use of shortcuts in which one or more of
eISSN1303-5150 www.neuroquantology.com
NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

the layers are skipped, and as a result the partial derivatives to determine the gradient of
training loss is lower; stacked layers are added the loss function as it is applied to the different
to the output; therefore, complex computations weights within the neural network. In this
are not required.ResNet18 shortcut connections manner, the loss function is minimized
omit two layers.The design used activation repeatedly until a termination criterion is met
functions described in Swish. or until it stops reducing.
Backpropagation and partial derivatives Chain Rule
Essentially, backpropagating is the It summarizes how a loss function, z, is
process of computing partial derivatives of the affected by some neural network parameter,
loss function as the weights (parameters) are let's say x and y, which are functions of the
changed in a network. To minimize the distance previous layer parameter t. We can consider
in the desired predictions, we use the chain layers f, g, and h to perform nonlinear
rule. Therefore, backpropagation involves using operations on input.
z=f(x,y)x=g(t)y=h(t) (1)
In calculus, suppose you want to express the gradient of z with respect to input z. Multivariable calculus
entails the following:

(2)
The values are often all less than 1, experimentally validated. Architecture with skip
regardless of the sign.Multiplication of partial connections skips some layers in the neural
derivatives is used in backpropagation to network and feeds the output from the
propagate gradients to earlier layers.While previous layer as input to the subsequent layers 7761
there is no strict mathematical theorem when it (instead of just the previous layer).
comes to multiplication with absolute values
less than 1, it gives the impression of training Skip Connections via Addition
stability. In the network, one can observe that It is however evident that for every
the gradient diminishes with every layer in successive layer that is backpropagated through
reverse in the network. the identity functions; the core idea is to just
Skip Connection add vectors. As a result, the gradient value in
Much convolutional architecture the earlier layers will be maintained by
currently use skip connections as a standard multiplying it by one. Stacking these skip
module. We provide an alternative gradient residual blocks together is the main idea behind
path (using backpropagation) by using a skip Residual Networks (Resenets). The gradient is
connection. This additional path is often preserved by using an identity function.
beneficial to the model convergence, as it is

eISSN1303-5150 www.neuroquantology.com
NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

7762

Fig-3 Skip Connection Architecture

Based on the loss function, we can calculate the partial derivative (gradient) of the residual
block as follows:

(3)
Early layers have been observed to extract lower semantic information from the input through
acquired features.
Activation Function
ReLU thresholds all negative weights to zero, while Swish propagates a small number of
negative weights. As deep neural networks become increasingly complex, this property becomes crucial
in the success of non-monotonic smooth activation functions, such as Swish.

eISSN1303-5150 www.neuroquantology.com
NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

7763

Fig 3- Architecture of proposed MResNetClassification Technique

Table 5 MResNet Layers details

MResNetAlgorithm

eISSN1303-5150 www.neuroquantology.com
 NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

Step 1: Insert the image (227x227x3).

Step 2:Convolutional layer 1(7x7size 64 filters are used). It generates a 112x112x64 pixel feature map.
Step 3: TheMax pooling (3x3 size 64filters are used). The output is a 56x5x64 pixel feature map.
Step 4: Convolutional layer 2(3x3size 64 filters are used). ). It generates a 56x56x64 pixel feature map.
Step 5: Convolutional layer 3(3x3size 64 filters are used). It generates a 56x56x64 pixel feature map.
Step 6: Convolutional layer 4(3x3size 64 filters are used). It generates a 56x56x64 pixel feature map.
Step 7: Convolutional layer 5(3x3size 64 filters are used). It generates a 56x56x64 pixel feature map.
Step 8: Convolutional layer 6(3x3size 128 filters are used). It generates a 28x28x128 pixel feature map.
Step 9: Convolutional layer 7(3x3size 128 filters are used). It generates a 28x28x128 pixel feature map.
Step 10: Convolutional layer 8(3x3size 128 filters are used). It generates a 28x28x128 pixel feature map.
Step 11: Convolutional layer 9(3x3size 128 filters are used).It generates a 28x28x128 pixel feature map.
Step 12: Convolutional layer 10(3x3size 256 filters are used). It generates a 14x14x256pixel feature map.
Step 13: Convolutional layer 11(3x3size 256 filters are used). It generates a 14x14x256 pixel feature map.
Step 14: Convolutional layer 12(3x3size 256 filters are used). It generates a 14x14x256 pixel feature map.
Step 15:.Convolutional layer 13(3x3size 256 filters are used). It generates a 14x14x256 pixel feature map.
Step 16: Convolutional layer 14(3x3size 512 filters are used). It generates a 7x7x512 pixel feature map.
Step 17: This layeris omitted from the Skip connections of MResNet.
Step 18: This layeris omitted from the Skip connections of MResNet.
Step 19:.Convolutional layer 17(3x3size 512 filters are used). It generates a 7x7x512 pixel feature map.
Step 20: The average pooling (7x7 size 512 filters are used). The output is a 1x1x512 pixel feature map.
Step 21: This layer is the output layer.It uses a Softmax activation function.
7764

Experimental Setup Evaluation Metrics

Dataset Description Metrics for evaluating statistical or
The training and assessment of our machine learning models are used in order to
approach were performed using a brain tumor determine their quality.To test a model, there
dataset with imitated acoustic features. This are many different evaluation metrics.For
dataset is available on Kaggle.The dataset different kinds of problems, different evaluation
consists of 3264 brain MRI images classified into metrics are used.
four groups: glioma, meningioma, pituitary, and
no tumor.
True positive + Truenegative
Accuracy= _________________________________________ (4)

True positive +True negative +False positive+ False negative

True positive
Precision = __________________________ (5)
True positive +False positive

True positive

eISSN1303-5150 www.neuroquantology.com
NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

Recall = __________________________ (6)

True positive +False negative

True negative
Specificity = __________________________ (7)
True negative +False positive

Precision x Recall
F1 Score = 2 x ________________________ (8)
Precision + Recall

Results LeNet, AlexNet, VGG Net-16, and ResNet-

Accuracy, precision, recall, Specificity, 18.The proposed(MResnet) approach achieves 7765
F1 score are used to evaluate performance. classification accuracy of 96.5 %.
Performance comparisons are performed on

Table 6 -Classification performance analysis

Algorithm Accuracy Precision Recall Specificity F1-Score
Lenet 84.48 89.66 81.25 88.48 85.25
Alexnet 87.66 93.1 84.38 92.31 88.52
VGG 16 89.66 96.55 84.85 96 90.32
ResNet 16 96.4 95.7 95 93 94.2
MResnet 96.5 92.3 95 93 94.12

100

95

90
Accuracy
85
Precision
80 Recall

75

70
Lenet Alexnet VGG 16 ResNet 16 MResnet

Fig 8- Classification Graphical Representation

Conclusion An accurate diagnosis of brain tumors

can positively influence cancer detection,
eISSN1303-5150 www.neuroquantology.com
NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

treatment, and outcome. This paper presents 4. Hareem Kibriya, Rashid Amin , Asma Hassan
high-fidelity classification architecture for the Alshehri, Momina Masood, Sultan S.
detection of brain tumors by MRI images. The Alshamrani , and Abdullah Alshehri, A Novel
depth and simplicity of CNN's processing and Effective Brain Tumor Classification
provide significant accuracy improvements in Model Using Deep Feature Fusion and
medical image datasets. Proposed method Famous Machine Learning Classifiers”,
MResNet used for by jumping layers in deep Computational Intelligence and
architectures, this skips some layer in the neural Neuroscience, pp.1-15, 2022.
network and passes the output of one layer as 5. Hassan Ali Khan, Wu Jue, Muhammad
an input to the next layer. It’s used to change Mushtaqand Muhammad Umer Mushtaq,
activation function relu to swish. It is proposed “Brain tumor classification in MRI image
to use deep CNNs based on MResNet to using convolutional neural network”, Math.
automatically classify brain tumor MR images Biosci. Eng,Vol 17, Issue no 5, pp.6203-
into three classes: gliomas, meningiomas, and 6216, 2020.
pituitary tumors. Further, experimental findings 6. K. Sarath Chandra, A. Sai Priya, S. Durga
show that the proposed technique is Maheshwari, Dr. B. Ramesh Naidu
significantly better in terms of overall accuracy “Detection of Brain Tumour by integration
precision, Recall, specificity, and F1-Score than of VGG-16 and CNN Model ” , International
the other four recent methods for brain tumor Journal of Creative Research Thoughts
classification. (IJCRT), Volume 8, Issue 7,pp. 298-304,
2020.
Reference 7. Krishnammal, P.M., Raja, S.S,
1. Alex Krizhevsky, A, Sutskever, Ilya Sutskever “Convolutional Neural Network based 7766
and Geoffrey E. Hinton, “Imagenet Image Classification and Detection of
classification with deep convolutional Abnormalities in MRI Brain Images”,
neural networks”Advances in neural International Conference on
information processing systems, volume 25, Communication and Signal Processing
pp.1-9, 2012. (ICCSP), IEEE, 4–6; pp. 0548–0553, 2019.
2. Chetana Srinivas, Nandini Prasad .K.S, 8. Mayanmosha Wahlang, Arnab Kumar Maji,
Mohammed Zakariah, Yousef Ajmi Alothaibi Goutam Saha, Prasun Chakrabarti, Michal
, Kamran Shaukat ,B. Partibane, and Halifa Jasinski, Zbigniew Leonowicz and Elzbieta
Awal “Deep Transfer Learning Approaches Jasinska, “Brain Magnetic Resonance
in Performance Analysis of Brain Tumor Imaging Classification Using Deep Learning
Classification Using MRI Images” Journal of Architectures with Gender and Age”
HealthcareEngineering Volume 2022, pp. 1- Sensors , volume 5,pp. 1-20, 2022.
17, 2022. 9. Modupe Odusami , Rytis Maskeliunas ,
3. Francisco Javier Díaz-Pernas, Mario Robertas Damaševicius, and Tomas
Martínez-Zarzuela, Míriam Antón- Krilavicius, “Analysis of Features of
Rodríguezand David González-Ortega “A Alzheimer’s Disease: Detection of Early
Deep Learning Approach for Brain Tumor Stage from Functional Brain Changes in
Classificationand Segmentation Using a Magnetic Resonance Images Using a
Multiscale Convolutional Neural Network” Finetuned ResNet18 Network” ,
Healthcare, Multidisciplinary Digital Diagnostics, Vol 11,Issue no 6, pp.1-16,
Publishing Institute,Vol. 9, Issue No. 2, pp. 2021.
1-14, 2021. 10. Modupe Odusami, Rytis Maskeliunas,
Robertas Damaševicius and Tomas
eISSN1303-5150 www.neuroquantology.com
NeuroQuantology| NOVEMBER 2022 | VOLUME 20 | ISSUE 15 |PAGE 7750-7767| DOI: 10.14704/NQ.2022.20.15.NQ88767
Mr.R.Vinayagamoorthy et al/ A Modified Resnet (MRestNet) based Classification for Brain Tumor Dataset

Krilavicius, “Analysis of Features of 17. Soumick Chatterjee, FarazAhmed Nizamani,

Alzheimer’s Disease: Detection of Early Andreas Nürnberger & Oliver Speck,
Stage from Functional Brain Changes in “Classifcation of brain tumours in MR
Magnetic Resonance Images Using a Fine images using deep spatiospatial models” ,
uned ResNet18 Network” , Diagnostics 11, Scientific Reports, 12(1), pp.1-11, 2022 .
Issue no. 6, 2021. 18. Sunita M. Kulkarni , Dr. G. Sundari, “ A
11. Ouiza Nait Belaid, Malik Loudini, Framework for Brain Tumor Segmentation
“Classification of Brain Tumor by and Classification using Deep Learning
Combination of Pre-Trained VGG16 Algorithm” , International Journal of
CNN”, Journal of Information Technology Advanced Computer Science and
Management, Volume 12 Issue No2, pp.13- Applications, Vol 11,Issue no. 8, pp. 374-
25, 2020. 382,2020.
12. P.Harish, S.Baskar, “MRI based detection 19. Tahia Tazin , Sraboni Sarker, Punit Gupta ,
and classification of brain tumor using Fozayel IbnAyaz, Sumaia Islam, Mohammad
enhanced faster R-CNN and Alex Net Monirujjaman Khan , Sami Bourouis , Sahar
model”, Materials Today: Proceedings,pp.1- Ahmed Idris,and HammamAlshazly “A
6, 2020. Robust and Novel Approach for Brain
13. Pallabi Sharma, ImayanmoshaWahlang, Tumor Classification Using Convolutional
SugataSanyaland Arnab Kumar Maji Neural Network”
“Classification of Brain MRI Using Deep Computational Intelligence and
Learning Techniques” Soft Computing: Neuroscience Volume 2021,pp.1-11, 2021.
Theories and Applications, Advances in 20. Yunzhe Xue, Yanan Yang, Fadi G. Farhat,
Intelligent Systems and Computing Frank Y. Shih, Olga Boukrina, A.M. Barrett,
7767
Springer, Volume 1053,pp. 559-569,2020. Jeffrey R. Binder, William W. Graves, and
14. Rayene Chelghoum, AmeurIkhlef Usman W. Roshan, “Brain tumor
,AminaHameurlaine , and SabirJacquir classification with tumor segmentations and
“Transfer Learning Using Convolutional a dual path residual convolutional neural
Neural Network Architectures for Brain network from MRI and pathology images.
Tumor Classification from MRI Images” IFIP In International MICCAI brain lesion
International Conference on Artificial workshop, Springer, pp. 360-367, 2019.
Intelligence Applications and Innovations. 21. Zahra Sobhaninia , Nader Karimi ,
Springer, pp. 189-200, 2020. PejmanKhadivi, Shadrokh Samavi1, “Medial
15. R.Vinayaga Moorthy, Residual Encoder Layers for Classification
Dr.R.Balasubramanian,“An Efficient Genetic Brain Tumors in Magnetic Resonance
Boruta(GenBoruta) Algorithm Based Images” arXiv, pp.1-4, 2020.
Feature Selection on Brain
Tumor Dataset” International Journal of
Mechanical Engineering, Vol.7 No.6, pp 693-
712, 2022.
16. Si-Yuan Lu, Suresh Chandra Satapathy, Shui-
Hua Wang and Yu-Dong Zhang, “PBTNet: A
New Computer-Aided Diagnosis System for
Detecting Primary Brain Tumors”, Frontiers
in Cell and Developmental Biology,pp.1-11,
2021.

eISSN1303-5150 www.neuroquantology.com