J. Inst. Eng. India Ser.

B
https://doi.org/10.1007/s40031-023-00916-w

ORIGINAL CONTRIBUTION

Deep Learning‑Based Robust Hybrid Approaches for Brain

Tumor Classification in Magnetic Resonance Images
S. Rama Krishna1 · Chittibabu Ravela2 ·
Srihari Varma Mantena3 · Mohammad Sirajuddin4 ·
G. Joel Sunny Deol5

Received: 29 May 2023 / Accepted: 28 August 2023

© The Institution of Engineers (India) 2023

Abstract Classification of brain images is a very chal- above are built using Python. The suggested models exhibit
lenging problem among the most helpful and commonly a high accuracy of 93.9% and 91.3% for a brain MRI dataset,
employed procedures in the medical system. Deep learn- respectively. Furthermore, the proposed models were com-
ing, a subset of artificial intelligence, has pioneered new pared and evaluated against other existing hybrid models in
techniques for automatic medical image investigation. The terms of dice score, Positive predictive value, False predic-
goal of this study was to create two deep learning-based tive value, precision, recall, specificity, and F1 score. As a
hybrid models to segment and classify brain MRI data. The result, the proposed deep-learning technique can assist doc-
first hybrid model is a combination of Fully Convolutional tors as well as radiologists in detecting brain tumors early.
Networks and Residual Networks for segmentation and clas-
sification. The second is a hybrid of SegNet for segmentation Keywords MRI · Brain tumor classification · Deep
and MobileNet for classification. The analysis is based on learning · Threshold-based segmentation
a labelled dataset containing images of glioma, meningi-
oma, pituitary, and no tumor. The hybrid models proposed Abbreviations
CNS Central Nervous System Cancers
MRI Magnetic Resonance Imaging
* S. Rama Krishna
[email protected] SPECT Simple Photon Emission Computed
Tomography
Chittibabu Ravela
[email protected] ReLU Rectified Linear Unit
FCN Fully Convolutional Networks
Srihari Varma Mantena
[email protected]
Mohammad Sirajuddin
[email protected] Introduction
G. Joel Sunny Deol
[email protected] A brain tumor is a condition marked by the growth of aber-
1 rant brain tissues and cells [1]. Every new cell displaces the
Department of Computer Science and Engineering, GITAM
(Deemed to Be University), Hyderabad, India one before it as cells divide and multiply in a predictable
2 order. Every new cell that forms during cell division and
Computer Science and Engineering, Koneru Lakshmaiah
Education Foundation, K L University, Vaddeswaram, India growth replace the one that came before it in a predictable
3 pattern. Some cells, on the other hand, are abnormal and
Department of Computer Science and Engineering, SRKR
Engineering College, Bhimavaram, India multiplied exponentially, producing considerable disruption
4 to brain function, and ultimately death. There are about 120
Department of Information Technology, Kallam
Haranadhareddy Institute of Technology, Dasarapalem, India different forms of brain Tumors and central nervous system
5 cancers (CNS). Moreover, it’s estimated that 9.6 million
Department of Computer Science and Engineering, Kallam
Haranadhareddy Institute of Technology, Dasarapalem, people lost their lives to cancer in 2018 [2]. Furthermore,
Andhra Pradesh, India

13
Vol.:(0123456789)
 J. Inst. Eng. India Ser. B

between 30 and 50% of those diagnosed with primary segmentation, and other areas, with promising results. This
tumors. One of the worst kinds is cancer is brain cancer. prompted the scientists to propose a deep learning network
According to the data [3] about 17,760 adults died in to diagnose illness categories based on MRI data.
2019 as a result of brain tumors. As per the American Can- Approaches based on deep learning, such as fully con-
cer Society, brain and CNS cancers will kill 18,600 people volutional networks (FCN) [12], U-net [19], ResNet [13],
and 3460 children under the age of 15 in the year 2021. The and Segnet [14, 15], were used previously, and they often
life expectancy for individuals with brain tumors was 36%, produce significantly accurate segmentation outputs than
whereas the 10 years rate of survival seems to be 31% [4]. the preceding two techniques. As a result, deep learning-
The absence of accurate information on tumor size based algorithms are used in this study to segment brain
caused by low-quality images of tumor areas makes early- tumor images. After segmentation, classification takes place.
stage brain tumor diagnosis challenging. Patients can be Various researchers have proposed various ways of classify-
effectively treated if the tumor is detected and treated at an ing brain tumors. However, hybrid models outperform all
initial phase of development. As a result, tumor treatment others in brain tumor categorization [20–23]. In this study,
is strongly reliant on rapid tumor diagnosis and classifica- two hybrid models for detecting brain tumors are proposed.
tion. To detect brain malignancies, many medical imaging These models are then compared to determine which one
techniques like Simple Photon Emission Computed Tomog- delivers the best accurate segmentation of brain tumors. Fol-
raphy (SPECT), Ultrasound, Positron Emission Tomogra- lowing segmentation, MRI image classification is carried
phy, X-ray, Magnetic Resonance Imaging (MRI), as well as out.
Computed Tomography (CT) scans are employed. Since it The dataset used in this study comprises the following
offers higher contrast images of tumors than other medical types: meningioma, pituitary, glioma, and no tumor. The
imaging technologies, one of the most often utilized medical dataset under consideration is probably biased since some
imaging techniques is MRI [5, 6]. brain disorders are more prevalent than others. To address
The information provided by MR imaging is particularly such issues, data augmentation is utilized, which balances
helpful in detecting tumors located in the brain [7]. Seg- the data and generates extra images from the present data via
menting the brain images with high accuracy from MRI rotations or transformations. Figure 1 shows various classes
is essential for identifying brain tumors. The manual MRI which are classified using the proposed models.
image segmentation is a tough process. This also necessi- The contributions can be listed as follows:
tates non-uniform segmentation, an unrepeatable assign-
ment, a large duration of time, and segmentation outcomes.
It varies from expert to expert at times, so the computer-
assisted tool is really useful in this case [8].
Numerous automated brain tumor segmentation tech-
niques were developed over the last few decades, and they
can be broadly classified into three types: Atlas registration-
grounded models [9, 10], Machine learning models [11], as
well as deep learning models [12–16] with automatic end-
to-end learning characteristics.
When the three approaches are compared, it is obvious
that the deep learning approaches have automated learning
features and are highly efficient for feature expression. It has
overcome the constraints of traditional methods for brain
tumor segmentation and classification and it has emerged
as the key research topic in the field of brain tumor analysis.
The identification of the disease kind with high accu-
racy may help patients receive a better therapy. The models
based on deep learning can handle difficulties by automati-
cally identifying input data characteristics as well as they are
adaptable to variations in the topic under investigation. Deep
learning models can find and study properties in unstudied
data patterns using inferred data, leading to good efficiency
also with minimal computer modeling. Deep learning tech-
nologies [17, 18] have also been progressively used in image Fig. 1  Sample brain tumor dataset for four classes: a meningioma, b
classification, target detection, object recognition, image Pituitary c no tumor d glioma

13
J. Inst. Eng. India Ser. B

• To propose two hybrid deep learning-based models be increased further by employing larger data sets and other
named FCN32 and Segnet for brain tumor segmentation deep-learning approaches.
and classification (glioma, meningioma, pituitary, and Wang et al. [19] described a new cross-level linked
no tumor). The first model (Model 1) is a combination of U-shaped networking model (CLCU-Net) for automated
fcn_32 for segmentation and resnet_50 for classification. and efficient brain tumor segmentation using MRI data. To
The second model (Model 2), which is a combination of properly exploit various scale data from brain tumors of dif-
Segnet for segmentation and Mobilenet for classification. ferent sizes, the CLCU-Net employs cross-level feature link-
• We utilize two models to split the human brain MRI data- ages. However, it has several drawbacks. To begin, whereas
set into training and testing data. MRIs are three-dimensional, the tests were conducted
• These two models were used to classify brain tumors as on two-dimensional images and the connection between
non-neoplastic, glioma, meningioma, and pituitary tumor slices is neglected. This could affect the performance of
using the well-known ResNet50 and MobileNet classi- segmentation.
fiers. A new 3D FCN technique for segmenting tumor cells
• To assess the proposed models using various parameters in MRI images was presented in [12]. More research is
such as accuracy, specificity, F1-score, loss, confusion required to investigate the usefulness of various normalized
matrix, ROC, dice score, precision, recall, PPV, and FPV. functions, such as batch normalization with changing batch
sizes, layer and group normalization, and instance and local
The rest of the article is structured as follows. response normalization.
Sect. "Related Works" addresses a recent study on iden- Shehab et al. [13] proposed an automated technique for
tifying brain tumors in MR images using various models. the segmentation of brain images to address the gradient
Sect. "Materials and Methods" explained the proposed issue in DNN. While comparing DNN, the proposed tech-
work’s methods. Sect. "Results" covered the assess- niques obtained high accuracy so this will make the process
ment matrices as well as the results of the chosen dataset. of training quick. While comparing other algorithms, the
Sect. "Conclusion" concludes this work. proposed method obtained a minimum computation running
time that is 3 times faster than such as CNN, Unet-res, and
FCN (U-Net).
Related Works Rehman et al. [14] presented BrainSeg-Net, which is a
model based on an encoder and decoder. This architecture
Shah et al. [1] built a powerful, EfficientNet-B0-model to has a block called the Feature Enhancer (FE) that extracts
identify brain tumor from MRI image. Image enhancement middle-level features from shallow layers and distributes
methods are employed to improve the quality of images them to dense layers.
by applying different filters. The number of data samples Segmentation by humans takes a lot of time as well
available for training the proposed model is expanded using as causing many errors, and also its efficacy is depend-
data augmentation techniques. Even though this research ent on doctors’ knowledge. Alqazzaz et al. [15] address
concentrated on five different convolutional models as well this issue by deploying a SegNet model for MRI modali-
as transfer learning methods in this work, more research is ties in 3-dimensional data sets to automatically segment
required. Time complexity also needs to be decreased. brain tumors and their sub-tumor components. The training
A hybrid technique utilizing brain MRI scans to identify step, on the other hand, is time-consuming, which could be
and categorize the cancer was used in this research work, regarded as a constraint.
and the BRATS database was used. Khairandish et al. [4]. Al-Saffar and Yildirim et al. (2020) [25] developed an
By utilizing the hybrid CNN with SVM algorithms, the automated approach for detecting and classifying brain
developed system attempts to categorize brain images as glioma stages. There are various restrictions that the
malignant or benign tumors. This work has to be improved designed methodology may suffer due to the selected fea-
because it does not consider the exact location of the tumor. tures. MI-ASVD (mutual information-accelerated singular
Sunil Babu and Vijayalakshmi [6] suggested an abnor- value decomposition) performs well on the data described
mal-to-normal translation generative adversarial network in this study, however different datasets could offer other
(ANTGAN) for converting a medical image with lesions issues and difficulties and the proposed approach may be
into a matching image with the lesion "removed." unsuitable for it. Deshpande et al. [26] aimed to increase
Asif et al. [24] developed a CNN technique based on the accuracy of brain tumor categorization by merging CNN
deep learning to identify the brain tumor from MRI. To approaches using a super-resolution procedure. The choice
obtain deep features from brain MRI, pre-trained models of the relevant blocks across frames in scan image series
like Xception, DenseNet121, InceptionResNetV2 as well as for fusing is aided by an unique Discrete Cosine Transform
NasNet Large are employed. The system’s performance can (DCT)-based fusion approach. This approach is known as

13
 J. Inst. Eng. India Ser. B

DCT-CNN-ResNet50. To demonstrate its effectiveness, this

model must be tested on a variety of datasets. The authors
of [27] constructed an image-processing framework that is
SVM based to identify brain tumors.
A quick introduction to various image segmentation
algorithms was provided in [28]. Threshold-based, Region
growth, Watershed, Edge detection, as well as K-mean clus-
tering approaches are the primary segmentation techniques
employed in the investigation. The presented image segmen-
tation algorithms are evaluated upon MRI-scanned images
that contain tumors. A new automatic recognition and clas-
sification method has been proposed [29]. A Brain tumor
images used enhanced K-means clustering methods during
the segmentation phase.

Materials and Methods

The proposed method for classifying brain tumors is

described in detail in this section. The proposed approach
is separated into two stages. We started by explaining the
experimental dataset that had been utilized for the brain
tumor segmentation and classification process. Second, we
discussed proposed deep learning approaches as well as
their design for identifying and classifying MRI images as Fig. 2  Overview of the proposed hybrid models
glioma, meningioma, pituitary, or no tumor in detail.
Both methods categorize the output into four classes such as
Dataset glioma, meningioma, no tumor, and pituitary.

The Magnetic Resonance Imaging (MRI) Dataset was used Pre‑processing

in this study [30]. The first openly accessible database of
brain MRI scan images was made available on the Kaggle According to Kang et al. [31], most brain MRI images have
Brain Tumor Detection 2020. This database comprises 7023 undesirable data in the type of noise, leading to a poor
MRI images classified as tumor as well as non-tumor (1500 classification result. To provide valuable information, it
for every class category). This research takes into account is required to reduce the noise and undesirable areas. The
4916 image files for the training set, 210 image files for cropping approach was used to find the extreme point; the
validation, and 1897 image files for testing. erosion and dilation operations were used to eliminate noise;
additional information about this method may be found in
Method Outline [31, 32]. All of the brain images in the MRI datasets had
different widths, heights, and sizes; all of the images were
This section gives an overview of the proposed models. This scaled to 512 × 512 to provide uniformity. Lastly, before
research uses deep learning techniques to improve the clas- giving the images to the model, they were all normalized/
sification accuracy of MR brain images. Figure 2 depicts the scaled.
overview of proposed hybrid models for classifying brain
tumors. Data Augmentation
The proposed framework model is divided into four
stages. The approach starts with a collection of MRI brain Data augmentation refers to a range of approaches for mas-
scan images, which are then pre-processed and augmented sively raising the amount of data by generating extra data
with data. The dataset is separated into three parts: the train- points with the help of available data. This is a rapid method
ing set, the test set, as well as validation set. This paper for increasing dimensionality and improving generalization
presents two hybrid models: Model 1 with FCN-32 seg- to formerly unknown samples through changing a small
mentation and ResNet-50 classification, and Model 2 with amount of data or generating samples with help of utilizing
SegNet for segmentation and MobileNet for classification. a deep network in the dataset. Several approaches can be

13
J. Inst. Eng. India Ser. B

employed for augmentation; nevertheless, for dataset size performed using upsampling layers to enhance segmenta-
improvement, this article used processes such as shifting tion outcomes. Based on this encoder and decoder out-
width and height, shear intensity, height shifting, bright- puts, every pixel present in the image is assigned to one
ness, and flipping in horizontal and vertical flips. This can of the available classes by the softmax function. Hence,
increase performance while preventing model overfitting. the output from the encoder-decoder model is considered
a feature representation for the input image, which can be
Model 1: FCN_32 with ResNet_50 known as pixel-wise feature maps.
Convolutions are performed on the input of each convo-
Because of their capacity to produce deep feature represen- lutional layer of the encoder to create a collection of fea-
tations in a pixel-wise manner, FCNs have now garnered ture maps. Assume the input x with size h × w consists of
great success. So, this work proposes a hybrid model named convolutional kernels with size y × z can be expressed as,
Model 1 with FCN-32 for segmenting brain tumor images. ( )
Initially, the input MRI image size is 512 × 512 which is pre- ∑
(1)
j j i,j i
O =f b + 𝜀 ∗x
processed and then data augmentation is made to improve i
the performance. Finally, classification is performed using
ResNet-50 after segmentation. This is shown in Fig. 3.Only where Oj represents output feature map, b and 𝜀 denote bias
locally connected layers, such as the convolution layer, pool- and kernel, respectively. f denotes the activation function
ing layer, as well as upsampling layer make up FCNs. In this which is rectified linear unit used in this work to speed up
situation, no thick layer is used. An FCN is made up of two the training process, and it is expressed as
basic parts: for pixel-wise object representation, an encoder
and decoder-based design, and a softmax layer is utilized for
ReLU(x) = max (0, x) (2)
pixel-wise assignment. The present work employs segmentation which is a
Using several convolutional and pooling layers, encod- threshold based. Various thresholds and maximum values
ing is performed on the input image, and then decoding is

Fig. 3  Model-1with FCN_32 and ResNet_50 for brain tumor classification

13
 J. Inst. Eng. India Ser. B

on the test input image to improve the useful information. The ResNet50 model does simple training and has several
Figure 6 shows segmentation examples after training on benefits because it can learn residuals directly from images
the dataset. This method contains the intensity histogram rather than image characteristics [33]. The following equa-
of all pixels in the image via masking. A threshold will tion [34] defines a basic residual block.
then be employed to split the image into items. A higher ( { })
threshold, on the other hand, can drastically limit the b = F a, Wa + a (3)
number of positive samples, resulting in overfitting. If the
where Wa denotes the weight present in the weight matrix.
threshold is set too low, the samples will be increasingly
In the above equation, the residual mapping which is to be
redundant. As a result, the segmentation network would
learned is denoted as Fig. 4 depicts the block diagram of
struggle for discriminating between positive and negative
ResNet50 architecture.
samples, decreasing the training effect. As a result, dur-
The segmented output is processed by the ResNet-50
ing training, we propose a multi-threshold segmentation.
architecture, which begins with a convolutional layer. It con-
Following each convolution layer comes a pooling layer
sists of 64 filters as well as a kernel with a size of 7 × 7. Then
(max-pooling). Pooling produces N feature maps that have
there exists a max pooling layer. It consists of a stride with
been subsampled using factors. While this encoder’s convo-
a value of 2. Following this, there are 3 convolutional layers
lution as well as pooling procedures aid in feature represen-
present with different kernel sizes and several filters. Then,
tation via retaining relatively strong activation, spatial data
these three layers are repeated, producing a total of 9 layers.
present in a receptive field is lost. It can be significant during
Then, the model comprises 3 convolutional layers with dif-
the image segmentation process. To overcome this problem,
ferent filters and kernel sizes. These layers are repeated four
at the encoder’s output, an upsampling layer-based decoder
times, which gives a total of 12 layers. Likewise, the convo-
is included. The transposed convolution, which adapts the
lutional layers are repeated which is illustrated in Fig. 3. In
feature map to the input size, is efficient and simple to imple-
the last stage, average pooling is implemented and finished
ment in the decoding step (Upsampling). Batch normaliza-
by a fully connected layer with 1000 nodes and a softmax
tion (BN) was utilized after every convolutional layer, allow-
function for generating 1 layer. As a result, the output will
ing for quicker as well as more robust network training.
be glioma, meningioma, pituitary, and no tumor.
ResNet‑50 for Classification
Model 2: SegNet with MobileNet
FCN-32 segmented output is sent to the ResNet-50 model
for classifying MRI brain images. ResNet-50 is a CNN-
Segmentation of brain images is a difficult job, with small-
based model which consists of 50 layers and also makes the
scale tumors posing the greatest challenge. The spatial and
network deep when compared to VGG16. Instead of layers,
location data was lost because of nonlinear transformations
this network utilizes layer mapping. This type of mapping is
as well as continuous convolutions, even though more con-
known as residual mapping. The ResNet network uses skip
textual characteristics were retrieved later in the encoder. By
links to prevent the gradient from vanishing.
overcoming this issue, an effective method for brain image
The input and output vector can be represented as a and
segmentation capable of detecting small-scale tumors may
b, where H(a) denotes a stacked nonlinear layers function.
be built. As a result, this work proposed hybrid model-2,
The residual function may be noted as F(a) = H(a) − a, so
which consists of SegNet for segmentation and MobileNet
H(a) = F(a) + a which yields a residual function. In this,
for classification. Figure 5 is an illustration of this. The
F(a) and a denotes stacked nonlinear layers. When a par-
input MRI image is pre-processed in the same way as in
ticular layer deteriorates the model’s efficiency, regulariza-
model 1, and then data augmentation is performed before
tion immediately bypasses that layer, tackling the vanishing
segmentation.
gradient problems.

Fig. 4  A basic block diagram representation of ResNet50 architecture

13
J. Inst. Eng. India Ser. B

Fig. 5  Model-2with SegNet and MobileNet for brain tumor classification

The SegNet network is made up of encoder-decoder to understand and apply. Depending on the number of
networks. The encoder layers are the same as the VGG16 thresholds utilized, threshold-based techniques are classified
network’s convolutional layers. The segmentation mask is into two groups. They are segmentation methods based on a
built by the decoder using pooling indices from the associ- global or local threshold. The image is initially transformed
ated encoder’s max-pooling. Pooling indices are reused in to a binary image for threshold-based segmentation. Accord-
SegNet, requiring less memory than other models such as ing to the image’s intensity, the light pixels in the dark back-
U-Net. ground were separated by choosing a certain threshold level.
To reduce complexity and reduce the number of encoder Likewise, pixels with values more than the threshold were
parameters, fully connected layers are removed. The encoder assigned to value one, and those with values less than the
network is made up of five encoders, each one having thir- threshold were regarded as zero values in the image.
teen convolutional layers which match the first thirteen lay- The fundamental idea behind this method is to choose a
ers of the VGG16 network [35]. Convolution is performed threshold level. While considering one threshold level, sepa-
by each encoder with a filter bank to generate a set of fea- rating the images into two sections is a difficult procedure.
ture maps for batch normalization and rectified linear unit As a result, sections in the image have varying intensities.
(ReLU) functions. As each succeeding application of the A Multiple threshold levels were chosen and tumor
pooling procedure improves the spatial context, the feature regions from MRI images are extracted. Algorithm 1 depicts
map’s spatial resolution drops. The decoder network has 13 the proposed model’s pseudo code.
convolutional layers since each encoder layer has a matching
decoder layer. A multiclass soft-max classifier is given the MobileNet for Classification
output from the final decoder layer and uses it to determine
the class probabilities for each individual pixel. The segmented output is fed into the MobileNet model,
The simplest image segmentation technique employed which classifies different forms of brain tumors. MobileNet,
in the SegNet model is threshold-based segmentation. It when compared to MobileNet V2 [36], seems to be a
divides the MRI images into areas based on intensity levels CNN-based image recognition model which is frequently
with the help of one or more thresholds. It is straightforward employed. The MobileNet design has the advantage of

13
 J. Inst. Eng. India Ser. B

requiring less computing effort when compared to the classic MobileNet employs 3 × 3 depthwise separable convolu-
CNN approach, which makes it suitable for use on portable tions. It needs 8 to 9 times smaller processing when com-
devices and Computers with lower processing abilities [37, pared to normal convolutions.
38]. MobileNet architecture is a basic framework that com- Algorithm 1 Pseudocode for proposed hybrid models.
prises convolution layers and it may be utilized to segregate
data dependent on two controllable features. These features Input: MRI brain images, number of epochs, learning rate, batch
size
can effectively flip between the accuracy and latency factors.
In regards to, network reduction size, the MobileNet strategy Output: Glioma, meningioma, pituitary, and no tumor
is beneficial. Begin
The MobileNet model has 28 layers, which include a con- Step 1: Read the original MRI brain images
volutional layer, a depthwise convolutional layer, a pooling Step 2: Perform pre-processing & data augmentation
layer, and a softmax activation function. The MobileNet Step 3: Perform Threshold-based segmentation
model is built on depthwise separable convolutions, a form Transform RGB image to grayscale image (Ig)
of factorized convolution that separates a basic convolution Fix initial threshold (T1) & final threshold value (T2)//Using
kind into depthwise and pointwise convolutions. Depthwise Eqs. (8) & (9)
and pointwise convolutions are the two layers that make up Threshold period (TP) = 1/255;
depthwise separable convolutions (DS) (refer to Fig. 4). To Let Selective Threshold (Ts) = T1;
apply a single filter, depthwise convolutions were done to Best value = −inf
every input channel. The output from the depthwise layer While (T1 ≤ T2)
is therefore linearly combined using a basic 1 × 1 convo- Convert Ig to a binary image (Ib)
lution. MobileNet utilizes batch normalization and ReLU for (every period = 0 in Ib) do
nonlinearities in these layers. After extracting features from Compute the respective intensity of pixel from Ig
images through convolutional layers, The generated feature end
maps are reduced using a layer called Global Average Pool- Calculate π = std(I);
ing. Finally, a fully connected layer with a softmax function Calculate µ = mean(I);
divides the outcome into four groups. Determine the objective function of Ib through cv
For standard convolution with stride one and padding, the Cv = π/µ;
output feature map is generated as [39]: While (cv > bv)
∑ bv = cv;
Fc,d,j = Ka,b,i,j .Ic+a−1,d+b−1,i (4) end
a,b,i Ts = Ts + Tp;
Step 4: Classify the segmented image using model 1
The following computing cost is included in standard
Step 5: Classify the segmented image using model 2
convolutions:
Step 6: Evaluate outputs from steps (4) & (5) with test and
SK .SK .A.B.SI .SI (5) validation data
end
In the above equation, SK denotes the spatial convolution
of a Kernel K; The following are the initial and final threshold settings
I, A, and B denote an input feature map, and the number used in algorithm 1 to segment brain images:
of input and output channels accordingly.
Depthwise convolution containing one filter per input ( ) (Intensity)
Initial threshold T1 = (8)
channel is represented as: 255
∑
Fc,d,j = Ka,b,i,j .Ic+a−1,d+b−1,i (6) ( ) mean(Intensity)
a,b,i Final threshold T2 = (9)
255
where K represents the depthwise convolutional kernel of Textural observations and evaluation may aid in the
size SK × SK × A; F represents the output feature map. diagnosis, classification of tumor types, and evalua-
Depthwise convolution contains the following computa- tion of therapeutic response. A few of the relevant fea-
tional cost: tures are described in the following:
Mean (M): It is calculated by adding up the pixel val-
SK .SK .A.SI .SI (7)
ues in the image and dividing the result by the total num-
ber of pixels in the image.

13
J. Inst. Eng. India Ser. B

Table 1  Hyperparameters of Parameters Value i7 CPU with 8 GB RAM, as well as 4 GB Nvidia GPU).
hybrid models This was done in Python with the machine learning-based
Learning rate 0.0001 Keras module. Python seems to be compatible using Tensor-
Learning decay rate 0.9 Flow and can be employed to build neural networks. This is
Epochs 100 advantageous both for CPU and GPU operations. For fine-
Optimizer Adam tuning hyper-parameters of the model, a network search as
Activation function ReLu well as the parameters for which the model performs better
Test batch size 20 on the given test data are employed.
The process of training a model includes selecting the
best hyperparameters. Table 1 displays the proposed model’s
hyperparameters.

Evaluation Metrics

The following measures were used to assess the experimen-

tal results: accuracy, specificity, F1-score, positive predictive
value, precision, sensitivity, as well as false predictive value.
True Positive (TP), True Negative (TN), False Positive (FP),
as well as False Negative (FN) classes were used to calculate
these measurements. Positive pixels that have been correctly
classified are referred to as TP. Positive pixels that have been
erroneously identified are referred to as FN. Negative pixels
that have been accurately identified are termed to be TN. FP
refers to the improperly recognized negative pixels.

Confusion Matrix

The test data are used to construct a confusion matrix, which

measure the efficiency of the proposed model. The confusion
matrix’s rows include true class data, whereas the columns
include forecasted class data. This matrix has four outcomes:
TP, TN, FP, as well as FN.
Accuracy: The accuracy is represented as a percentage of
Fig. 6  Confusion matrix a Model-1 b Model-2 the total amount of accurately labeled pixels concerning the
entire range of pixels in the images. It examines the image’s
entire correct pixels.
Standard Deviation (SD): It seems to be the second central
TP + TN
moment representing a population’s probability distribution Accuracy = (10)
TP + TN + FP + FN
that may be used to assess inhomogeneity.
Entropy (E): It is used for quantifying the randomness pre- The confusion matrix for Model-1 and Model-2 are
sent in a textural image. shown in Fig. 6a and b. It signifies the findings of an inquiry
Skewness: It seems to be a measurement of symmetry. on an MRI brain imaging dataset for the classes, meningi-
Kurtosis: It can be defined as a parameter that indicates oma, pituitary, glioma as well as no tumor.
the shape of the probability distribution of a random variable. The precision can be measured by the ratio of precisely
Correlation: Correlation represents the spatial interde- labeled Positive samples to the entire number of Positive
pendence between pixels. samples classified (either correctly or incorrectly). The sen-
sitivity/recall of a classifier may be defined as the proportion
of how much was accurately recognized as positive to how
Results many of those were genuinely positive. The specificity of
a classifier can be defined as the proportion of the amount
The suggested approach is assessed with a public image that was correctly identified as negative to the genuinely
dataset (MRI Dataset) and performed on an HP laptop (Core negative amount. The F1 score is regarded to be a stronger

13
 J. Inst. Eng. India Ser. B

predictor of classifier performance than the standard accu- Similarly, PPV demonstrates the level of image classifi-
racy measure. cation accuracy. When models 1 and 2 were compared, the
proposed model 2 took the beneficial step shown in Table 2,
TP
Precision = (11) and it progressed at quite a faster rate. Other models, such
TP + FP
as [4, 20], and [18], do not outperform the proposed models.
In addition to the proposed models with a specified
TP assessment parameter, as shown in Table 3, other features
Sensitivity∕recall = (12)
FN + TP of the hybrid model, including FPV and PPV, provided sig-
nificantly better results than other current hybrid models.
TN Moreover, the proposed work does have a high rate as well
specificity = (13)
FP + TN as a higher classification result.
The Dice score [14] is utilized to determine the level of
F1 − score =
2 ∗ Precision ∗ Recall
(14) similarity between two sets, namely M and N, and therefore
Precision + Recall is stated as,
Table 2 compares the performance of proposed models 2 × |M ∩ N|
based on precision, recall, F1-score, and specificity. It is Dice score = (17)
|M| + |N|
obvious that Model 1 exhibit better precision of 96% and
F1-score of 94%, which are 1% and 1% higher when com- Here |M| and |N| represent the cardinalities for sets M and
pared to Model 2. Furthermore, Model 1 has a greater recall N, accordingly.
of 94% and a higher specificity of 98.7%, which are 2% and The dice score varies among the range 0 and 1, with 1
0.2% higher than Model 2. On an MRI dataset, the suggested representing the closest match between forecasted and actual
models Model 1 and Model 2 outperform other existing values. As seen in Table 2, both proposed models are closer
models including Hybrid deep autoencoder with a Bayesian to a value of one, indicating that they perform well during
fuzzy clustering-based segmentation [18], Google-Net with the segmentation process. When it comes to segmentation,
a soft-max classifier [20], and Hybrid CNN with SVM [4]. model 2 outperforms model 1. The predicted mask identi-
Positive Predictive Value: The TP also displays the like- fies the presence and location of the tumor, demonstrating
lihood that a given set of pixels was accurately classified. the proposed method’s great diagnostic usefulness. The dice
score suggests that the suggested models outperform cur-
TP
Positive predictive value = (15) rent models including a Hybrid deep autoencoder with a
TP + FP
False Predictive Value: In image processing, TN repre-
sents the probability pixel detection that theoretically indi-
Table 3  Results of various features used for segmentation
cates normal but it is recognized as an abnormal character-
istic whenever accuracy or TP are inadequate. Images Mean Standard Skewness Kurtosis Correlation
deviation
FP
False predictive value = (16) Fig. 6a 7.88 33.99 0.00651 1.79023E-06 0.998
FP + TP
Fig. 6b 9.90 32.19 0.05309 1.2560E-06 0.988
But, FPV is another crucial element for determining Fig. 6c 11.81 39.51 0.00456 2.89150E-06 0.987
whether or not an activity is being performed correctly. Fig. 6d 6.62 29.91 0.04363 2.2398E-06 0.998
Table 2 illustrates that Model 2 has a superior result because Fig. 6e 19.40 85.69 0.01017 2.4501E-06 0.978
it has completed this process with greater accuracy. Fig. 6f 3.93 59.23 0.0357 3.9120E-06 0.977

Table 2  Performance comparison of two models

Models Precision Recall F1 score Specificity Positive pre- False predic- Dice score
dictive rate tive rate

Model 1 0.96 0.94 0.94 0.987 0.79 0.83 0.774

Model 2 0.95 0.92 0.93 0.985 0.87 0.86 0.76
Hybrid CNN with SVM [4] 0.94 0.90 0.91 0.972 0.75 0.82 0.64
Hybrid deep autoencoder [18] 0.93 0.89 0.90 0.975 0.71 0.79 0.69
Google-Net with a softmax classifier [20] 0.87 0.90 0.88 0.961 0.69 0.81 0.67

13
J. Inst. Eng. India Ser. B

Bayesian fuzzy clustering-based segmentation, Google-Net Model 1 attained the highest validation and training accu-
with a soft-max classifier, and Hybrid CNN with SVM dur- racy when compared to Model 2b. This demonstrates the
ing segmentation. model’s accuracy for epochs = 100. Model 1 obtains a vali-
The threshold-based segmentation strategy is straight- dation accuracy of 93.9%, which is 2.6% greater than Model
forward and takes less time to process. As a consequence, 2.
depending on the intensity histogram of all pixels present in
the image, this research utilized threshold relied on segmen- Training and Validation Loss
tation. A threshold is then employed to split the image into
sections. In segmentation, for example, various threshold One of the crucial parts of neural networks is the loss func-
values ranging from 0 to 512 are generated for image pixels. tion, which measures model prediction error. The model
As a consequence, all pixels with values below or equiva- is trained using train data. Additionally, validation loss is
lent to the threshold can be allocated 0 as well as all pixels employed to evaluate how well a deep learning strategy
having values greater than the threshold can be allocated performs on the validation set for brain image categoriza-
512. The example output MRI images following training on tion. The loss calculated at both of these stages is known
the dataset are shown in Fig. 7. Figure 7a and b denotes a as the model loss. Figure 9a and b show the model loss for
Glioma tumor; Fig. 7c and d denotes a meningioma tumor; epochs = 100 on the training and validation datasets for
Fig. 7e and f denotes a pituitary tumor, respectively. Model 1 and Model 2, accordingly. At epoch 40, both model
Table 3 shows the results obtained for various features training and validation loss are nearly the same and at epoch
like mean, skewness, standard deviation, kurtosis, and 100, Model 1 loss is comparatively low when compared to
correlation. Model 2. This shows Model 1 has low training and vali-
dation loss. The MRI images are pre-processed to remove
Training and Validation Accuracy noise before passing through the models. Also, the Adam
optimizer in addition helps both models for optimizing the
Figure 8a and b depict the suggested models’ accuracy train- neural networks during a training. The training loss gradu-
ing and validation dataset. Figure 8a demonstrates that ally decreases in both models over the 100 epochs.

Fig. 7  Examples of segmen-

tation after training on the
dataset: a and b Glioma tumor;
c and d Meningioma tumor; e
and f pituitary tumor

13
 J. Inst. Eng. India Ser. B

Fig. 9  Training and validation loss a Model 1 b Model 2

Fig. 8  Training and validation accuracy a Model 1 b Model 2

from MRI scan images. Model 2 in Fig. 10b outperforms

As illustrated in Fig. 9a, the loss for Model 1 appears to Model 1 in Fig. 9a because the curve is closer to the left
be stabilized near the end. The ResNet model (Model 1) corner, indicating a greater true positive rate and accurate
is built upon shortcut connections that perform personality model performance.
mapping, as well as the output is connected to have stacked The Model 1 and Model 2 classifiers in this study had the
layers output with no further parameters. It indicates that greatest accuracies, with 98.45% and 98.85%, respectively,
layers present in the ResNet will try to acquire residual input when we used the test data itself. In Table 4, the accuracy of
as well as output. The vanishing gradient problem is handled the suggested model is compared to the findings of various
by the ResNet model’s shortcut. recent studies. The table shows that our concept is widely
connections. used, successful, and original.

ROC
Conclusion
The Receiver Operator Characteristic (ROC) curve is an
important metric to be considered during classification and In this study, a novel and effective technique for two
identification tasks. The assessment of specificity as well as hybrid models based on deep learning on the MRI dataset
sensitivity is critical for measuring the performance of the is presented for segmentation and classification of brain
proposed technique. The sensitivity can be calculated using tumors into glioma, meningioma, no-tumor, and pituitary
the TP and FN values, while the specificity will be calcu- classes. The initial processing steps were carried out on
lated with the help of TN and FP rates. While plotting the TP MRI images, which were subsequently segmented using a
rate and FP rate, the ROC curve appears. The ROC curve in threshold-based approach. The labeled segmented features
Fig. 9 were closer than the upper left corner, indicating that are then passed to the classifier, which uses them to clas-
the proposed approach accurately classifies tumor categories sify MRI images. According to the experimental results,

13
J. Inst. Eng. India Ser. B

The most crucial method for MRI-based brain tumor clas-

sification is presented in this paper.
One drawback of the present work is that the suggested
model was not assessed on a sufficiently large dataset.
Because of this, it will be important to assess the model’s
performance using big data in the future. Another drawback
is the lack of MRI testing of real patients in Bangladesh
for the suggested procedure. We will seek to establish MRI
at several hospitals and diagnostic facilities in Bangladesh
in the future, expanding on our effort to create a real-time
medical diagnostic system. We think our approach is ideal
for the medical industry, particularly for offering very pre-
cise and creative solutions for a more precise detection of
brain tumors.

Funding No funding received by any government or private concern.

Data Availability Data sharing not applicable to this article as no

datasets were generated or analyzed during the current study.

Declarations

Conflict of interest The authors declare that they have no competing

interests.

Research Involving Human Participants and/or Animals This arti-

cle does not contain any studies involving Human Participants and/or
Animals performed by any of the authors.

Fig. 10  ROC curve a Model 1 b Model 2 Informed Consent Not applicable.

Table 4  Comparison with other existing findings

Ref Classifiers Database Accuracy (%) References
[13] AlexNet BRATS2015 61.20 1. H.A. Shah, F. Saeed, S. Yun, J.-H. Park, A. Paul, J.-M. Kang, A
VGG16 97.20 robust approach for brain tumor detection in magnetic resonance
[22] CNN Nanfang Hospital, 96.3 images using finetuned EfficientNet. IEEE Access 10, 65426–
General Hospi- 65438 (2022). https://​doi.​org/​10.​1109/​ACCESS.​2022.​31841​13
tal, China 2. P.M. Shak, T.E.E. Tobely, H. Al-Feel, G. Manogaran, S. Baskar,
[23] CNN Kaggle 96.2 Neural network-based brain tumor detection using wireless infra-
red imaging sensor. IEEE Access 7, 5577–5588 (2019)
[30] DCNN M/s. Devaki Scan 95.5 3. F. Islami, C.E. Guerra, A. Minihan, K.R. Yabroff, S.A. Fedewa,
MDCNN Centre, India 97.4 K. Sloan et al., American cancer society’s report on the status of
[Proposed] Model 1 MRI 98.45 cancer disparities in the United States 2021. CA Cancer J. Clin.
Model 2 98.85 72(2), 112–143 (2022)
4. M. Khairandish, M. Sharma, V. Jain, J. Chatterjee, N. Jhanjhi, A
hybrid CNN-SVM threshold segmentation approach for tumor
detection and classification of MRI brain images. IRBM 43(4),
290–299 (2022). https://​doi.​org/​10.​1016/j.​irbm.​2021.​06.​003
5. S. Ahmad, P.K. Choudhury, On the performance of deep transfer
a hybrid Model 1 attained a maximum accuracy of 93.9%, learning networks for brain tumor detection using MR images.
which is 2.6% higher than Model 2.These models also IEEE Access 10, 59099–59114 (2022). https://​doi.​org/​10.​1109/​
were evaluated to existing techniques, which revealed that ACCESS.​2022.​31793​76
the proposed models outperform and provide optimal out- 6. M. Sunil Babu, V. Vijayalakshmi, An effective approach for sub-
acute ischemic stroke lesion segmentation by adopting meta-heu-
comes in terms of precision, recall, F1-score, Specificity, ristics feature selection technique along with hybrid naive Bayes
dice score, PPV, and FPV. In addition, the accuracy, loss, and sample-weighted random forest classification. Sens. Imaging
and ROC of the proposed hybrid models were compared. 20(1), 1–24 (2019). https://​doi.​org/​10.​1007/​s11220-​019-​0230-6

13
 J. Inst. Eng. India Ser. B

7. S. Banerjee, S. Mitra, B.U. Shankar, Automated 3D segmenta- multi-parametric MR scans. Open Comput. Sci. 12(1), 211–226
tion of brain tumors using visual saliency. Inf Sci 424, 337–353 (2022). https://​doi.​org/​10.​1515/​comp-​2022-​0242
(2018) 23. K.C.T. Swamy, V.V. Kishore, S.T. Ahmed and M.A. Farida,
8. A. Mukaram, C. Murthy, M.Z. Kurian, An automatic brain tumor Investigation of GPS-TEC inconsistency and correlation with
detection. Segmentation and classification using MRI image. Int. SSN, solar flux (F10.7 cm) and Ap-index during low and high
J. Electron. Electr. Comput. Syst. 6(5), 54–65 (2017) solar activity periods (2008 and 2014) over Indian equatorial low
9. S. Doyle,F. Vasseur, M. Dojat, F. Forbes, Fully automatic brain latitude region: in 2021 International Conference on Intelligent
tumor segmentation from multiple MR sequences using hidden Technologies (CONIT), Hubli, India, pp. 1–9. https://​doi.​org/​10.​
Markov fields and variational EM, in: Proceedings NCI-MICCAI 1109/​CONIT​51480.​2021.​94982​92
BraTS, (2013), pp. 18–22 24. S. Asif, W. Yi, Q.U. Ain, J. Hou, T. Yi, J. Si, Improving effective-
10. V. Kalpana, V. Vijaya Kishore, R.V.S. Satyanarayana, MRI and ness of different deep transfer learning-based models for detecting
SPECT brain image analysis using image fusion, in Mobile Radio brain tumors from MR images. IEEE Access 10, 34716–34730
Communications and 5G Networks Lecture Notes in Networks (2022). https://​doi.​org/​10.​1109/​ACCESS.​2022.​31533​06
and Systems. ed. by N. Marriwala, C. Tripathi, S. Jain, D. Kumar 25. V. Vijaya Kishore, V. Kalpana, ROI segmentation and detection of
(Springer, Singapore, 2023) neoplasm based on morphology using segmentation operators, in
11. V. Vijaya Kishore, V. Kalpana, Effect of noise on segmentation Emerging Trends in Electrical, Communications, and Information
evaluation parameters, in Soft Computing: Theories and Applica- Technologies. Lecture Notes in Electrical Engineering. ed. by T.
tions, Advances in Intelligent Systems and Computing. ed. by M. Hitendra Sarma, V. Sankar, R. Shaik (Springer, Singapore, 2020)
Pant, T. Kumar Sharma, R. Arya, B. Sahana, H. Zolfagharinia 26. A. Deshpande, V.V. Estrela, P. Patavardhan, The DCT-CNN-
(Springer, Singapore, 2020) ResNet50 architecture to classify brain tumors with super-reso-
12. V. Vijaya Kishore and R.V.S. Satyanarayana, Performance evalu- lution, convolutional neural network, and the ResNet50. Neurosci.
ation of edge detectors–morphology based ROI segmentation and Inform. 1(4), 100013 (2021). https://d​ oi.o​ rg/1​ 0.1​ 016/j.n​ euri.2​ 021.​
nodule detection from DICOM lung images in the noisy environ- 100013
ment. in 2013 3rd IEEE International Advance Computing Con- 27. V. Kalpana, V. Vijaya Kishore, K. Praveena, A common frame-
ference (IACC), Ghaziabad, India, 2013, pp. 1131-1137, https://​ work for the extraction of ILD patterns from CT image, in Emerg-
doi.​org/​10.​1109/​IAdCC.​2013.​65143​86 ing Trends in Electrical, Communications, and Information
13. L.H. Shehab, O.M. Fahmy, S.M. Gasser, M.S. El-Mahallawy, An Technologies. Lecture Notes in Electrical Engineering. ed. by T.
efficient brain tumor image segmentation based on deep residual Hitendra Sarma, V. Sankar, R. Shaik (Springer, Singapore, 2020)
networks (ResNets). J. King Saud Univ. Eng. Sci. 33(6), 404–412 28. V. Pooja, M.K. Kumar, K. Kamalesh, Comparative analysis of
(2021). https://​doi.​org/​10.​1016/j.​jksues.​2020.​06.​001 segmentation techniques on MRI brain tumor images. Mater.
14. M.U. Rehman, S. Cho, J. Kim, K.T. Chong, BrainSeg-Net: brain Today Proc. 47, 109–114 (2021). https://​doi.​org/​10.​1016/j.​matpr.​
tumor MR image segmentation via enhanced encoder-decoder 2021.​03.​723
network. Diagnostics (Basel) 11(2), 169 (2021). https://​doi.​org/​ 29. S. Dhana Sekaran, N.H.A. Rufus, D. Anand, R.S. Rama, A.
10.​3390/​diagn​ostic​s1102​0169.​PMID:​33504​047;​PMCID:​PMC79​ Kumar, A.S. Vigneshwar, Evolutionary optimization with deep
11842 transfer learning for content based image retrieval in cloud envi-
15. S. Alqazzaz, X. Sun, X. Yang et al., Automated brain tumor ronment. 2022 International Conference on Augmented Intelli-
segmentation on multi-modal MR image using SegNet. Com- gence and Sustainable Systems (ICAISS), Trichy, India, 2022, pp.
put. Vis. Media 5, 209–219 (2019). https://​doi.​org/​10.​1007/​ 826–831, https://​doi.​org/​10.​1109/​ICAIS​S55157.​2022.​10011​122
s41095-​019-​0139-y 30. https://​www.​kaggle.​com/​datas​ets/​sarta​jbhuv​aji/​brain-​tumor-​class​
16. N. Noreen, S. Palaniappan, A. Qayyum, I. Ahmad, M. Imran, M. ifica​tion-​mri
Shoaib, A deep learning model based on concatenation approach 31. M. Annamalai, P. Muthiah, An early prediction of tumor in heart
for the diagnosis of brain tumor. IEEE Access 8, 55135–55144 by cardiac masses classification in echocardiogram images using
(2020). https://​doi.​org/​10.​1109/​ACCESS.​2020.​29786​29 robust back propagation neural network classifier. Braz. Arch.
17. A. Manikandan, M.P. Bala, Intracardiac mass detection and clas- Biol. Technol. 65, e22210316 (2022). https://​doi.​org/​10.​1590/​
sification using double convolutional neural network classifier. J. 1678-​4324-​20222​10316
Eng. Res. 11(2A), 272–280 (2023). https://​doi.​org/​10.​36909/​jer.​ 32. R.C. Karpagalakshmi, D. Tensing, A.M. Kalpana, Image localiza-
12237 tion using deformable model and its application in health infor-
18. P. Siva Raja, A.V. Rani, Brain tumor classification using a hybrid matics. J. Med. Imaging Health Inf. 6(8), 1972–1976 (2016).
deep autoencoder with Bayesian fuzzy clustering-based segmenta- https://​doi.​org/​10.​1166/​jmihi.​2016.​1959
tion approach. Biocybern. Biomed. Eng. 40(1), 440–453 (2020). 33. K. Sheikdavood, P. Surendar, A. Manikandan, Certain investiga-
https://​doi.​org/​10.​1016/j.​bbe.​2020.​01.​006 tion on latent fingerprint improvement through multi-scale patch
19. Y.L. Wang, Z.J. Zhao, S.Y. Hu, F.L. Chang, CLCU-Net: cross- based sparse representation. Indian J. Eng. 13(31), 59–64 (2016)
level connected U-shaped network with selective feature aggre- 34. Y. Zhang, J. Li, S. Wei, F. Zhou, D. Li, Heartbeats classification
gation attention module for brain tumor segmentation. Comput. using hybrid time-frequency analysis and transfer learning based
Methods Programs Biomed. 207, 106154 (2021). https://​doi.​org/​ on ResNet. IEEE J. Biomed. Health Inform. 25(11), 4175–4184
10.​1016/j.​cmpb.​2021.​106154 (2021). https://​doi.​org/​10.​1109/​JBHI.​2021.​30853​18
20. M. Rasool, N.A. Ismail, W. Boulila, A. Ammar, H. Samma, 35. K. Namrata, R.C. Karpagalakshmi, S.S. Manikandasaran, Imple-
W.M.S. Yafooz, A.-H.M. Emara, A hybrid deep learning model mentation of novel technique for image watermarking using
for brain tumor classification. Entropy (Basel, Switzerland) 24(6), 2D-DCT. Int. J. Pure Appl. Math. 117(16), 221–226 (2017)
799 (2022). https://​doi.​org/​10.​3390/​e2406​0799 36. A.R. Venmathi, S. David, E. Govinda, K. Ganapriya, R. Dhanapal,
21. N. Cinar, A. Ozcan, M. Kaya, A hybrid DenseNet121-UNet model A. Manikandan, An automatic brain tumors detection and clas-
for brain tumor segmentation from MR images. Biomed. Signal sification using deep convolutional neural network with VGG-19.
Process. Control 76, 103647 (2022). https://​doi.​org/​10.​1016/j.​ 2023 2nd International Conference on Advancements in Electri-
bspc.​2022.​103647 cal, Electronics, Communication, Computing and Automation
22. S. Das, S. Bose, G.K. Nayak, S. Saxena, Deep learning- (ICAECA), Coimbatore, India, 2023, pp. 1–5, https://​doi.​org/​10.​
based ensemble model for brain tumor segmentation using 1109/​ICAEC​A56562.​2023.​10200​949

13
J. Inst. Eng. India Ser. B

37. R. Ali, A. Manikandan, J. Xu, A novel framework of adaptive Publisher’s Note Springer Nature remains neutral with regard to
fuzzy-GLCM segmentation and fuzzy with capsules network jurisdictional claims in published maps and institutional affiliations.
(F-CapsNet) classification. Neural. Comput. Appl. (2023). https://​
doi.​org/​10.​1007/​s00521-​023-​08666-y Springer Nature or its licensor (e.g. a society or other partner) holds
38. D. Balamurugan, S.S. Aravinth, P.C.S. Reddy, A. Rupani, A. exclusive rights to this article under a publishing agreement with the
Manikandan, Multiview objects recognition using deep learning- author(s) or other rightsholder(s); author self-archiving of the accepted
based Wrap-CNN with voting scheme. Neural Process. Lett. 54, manuscript version of this article is solely governed by the terms of
1–27 (2022). https://​doi.​org/​10.​1007/​s11063-​021-​10679-4 such publishing agreement and applicable law.
39. A.G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T.
Weyand, M. Andreetto, and H. Adam, MobileNets: efficient con-
volutional neural networks for mobile vision applications (2017).
arXiv [cs.CV]. http://​arxiv.​org/​abs/​1704.​04861

13