Hybrid Deep Convolutional Neural Network

Srinivasan et al. BMC Medical Imaging (2024) 24:21
https://doi.org/10.1186/s12880-024-01195-7
RESEARCH Open Access
©The Author(s) 2024. Open AccessThis article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made.The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.To view a copy of this
licence, visit http://creativecommons.org/licenses/by/4.0/.The Creative Commons Public Domain Dedication waiver (http://creativecom-
mons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
BMC Medical Imaging
A hybrid deep CNN model for brain tumor
image multi‑classification
Saravanan Srinivasan1
, Divya Francis2
, Sandeep Kumar Mathivanan3
, Hariharan Rajadurai4
,
Basu Dev Shivahare3
and Mohd Asif Shah5,6,7*
Abstract
The current approach to diagnosing and classifying brain tumors relies on the histological evaluation of biopsy
samples, which is invasive, time-consuming, and susceptible to manual errors. These limitations underscore the press-
ing need for a fully automated, deep-learning-based multi-classification system for brain malignancies. This article
aims to leverage a deep convolutional neural network (CNN) to enhance early detection and presents three distinct
CNN models designed for different types of classification tasks. The first CNN model achieves an impressive detection
accuracy of 99.53% for brain tumors. The second CNN model, with an accuracy of 93.81%, proficiently categorizes
brain tumors into five distinct types: normal, glioma, meningioma, pituitary, and metastatic. Furthermore, the third
CNN model demonstrates an accuracy of 98.56% in accurately classifying brain tumors into their different grades. To
ensure optimal performance, a grid search optimization approach is employed to automatically fine-tune all the rele-
vant hyperparameters of the CNN models. The utilization of large, publicly accessible clinical datasets results in robust
and reliable classification outcomes. This article conducts a comprehensive comparison of the proposed models
against classical models, such as AlexNet, DenseNet121, ResNet-101, VGG-19, and GoogleNet, reaffirming the superi-
ority of the deep CNN-based approach in advancing the field of brain tumor classification and early detection.
Keywords Brain tumor grading, Hybrid deep learning, Hybrid convolutional neural network, Grid search,
Hyperparameters
Introduction
Brain tumors stand as one of the leading causes of
death in the modern world. These tumors can manifest
in various regions of the brain, often remaining asymp-
tomatic until later stages of life. Symptoms of brain
disease encompass a wide array of issues, including
personality changes, memory difficulties, communica-
tion impairments, hearing or speech challenges, chronic
migraines, and even vision loss [1]. Notable examples of
brain tumors include meningiomas, gliomas, pituitary
adenomas, and acoustic neuromas. According to medi-
cal observations, meningiomas, gliomas, and pituitary
tumors account for approximately 15%, 45%, and 15% of
all brain tumors, respectively. A brain tumor can have
long-lasting psychological effects on the patient. These
tumors originate from primary abnormalities in the brain
*Correspondence:
Mohd Asif Shah
drmohdasifshah@kdu.edu.et
1
Department of Computer Science and Engineering, Vel Tech Rangarajan
Dr.Sagunthala R&D Institute of Science and Technology, Chennai 600062,
India
2
Department of Electronics and Communication Engineering, PSNA
College of Engineering and Technology, Dindigul 624622, India
3
School of Computing Science and Engineering, Galgotias University,
Greater Noida 203201, India
4
School of Computing Science and Engineering, VIT Bhopal University,
Bhopal–Indore Highway Kothrikalan, Sehore 466114, India
5
Department of Economics, Kabridahar University, Po Box 250, Kebri
Dehar, Ethiopia
6
Centre of Research Impact and Outcome, Chitkara University Institute
of Engineering and Technology, Chitkara University, Rajpura, Punjab,
140401, India
7
Division of Research and Development, Lovely Professional University,
Phagwara, Punjab, 144001, India

of 21
or central spine tissue that disrupt normal brain func-
tion. Brain tumors are classified into two main categories:
benign and malignant. Benign tumors grow slowly and
are non-cancerous; they are relatively rare and do not
metastasize. In contrast, malignant brain tumors con-
tain cancerous cells, typically originating in one region
of the brain before swiftly spreading to other areas of
the brain and spinal cord [2]. Malignant tumors pose a
significant health risk. The World Health Organization
(WHO) classifies brain tumors into four grades based
on their behavior within the brain: grades 1 and 2 are
considered low-grade or benign tumors, while grades
3 and 4 are categorized as high-grade or malignant
tumors. Several diagnostic methods, such as CT scan-
ning and EEG, are available for detecting brain tumors,
but magnetic resonance imaging (MRI) is the most reli-
able and widely utilized. MRI generates detailed internal
images of the body’s organs by employing strong mag-
netic fields and radio waves [3]. Essentially, CT or MRI
scans can distinguish the affected brain region due to
the tumor from the healthy tissue. Biopsies, clinical tests
that extract brain cells, can be conducted as a prelude to
cerebral surgery. Precision is paramount in measuring
tumor cells or arriving at accurate diagnoses. The emer-
gence of machine learning (ML) presents an opportunity
to assist radiologists in furnishing precise disease status
information [4]. The proliferation of novel technolo-
gies, particularly artificial intelligence and ML, has left
an indelible mark on the medical field, equipping vari-
ous medical departments, including medical imaging,
with indispensable tools to enhance their operations. As
MRI images are processed to aid radiologists in decision
making, a diverse array of automated learning strategies
is employed for classification and segmentation pur-
poses. While supervised methods for classifying brain
tumors hold immense promise, they demand specialized
expertise to optimize the feature extraction and selection
techniques [5]. In navigating and analyzing vast datasets,
expert medical professionals benefit from the support
of machine assistance. Furthermore, the failure to accu-
rately identify life-threatening tumors could potentially
result in treatment delays for patients. The utilization of
deep-learning (DL) techniques in detecting brain tumors
and extracting meaningful insights from data patterns
has a longstanding history. DL’s capability to classify and
model brain cancers is widely recognized [6]. Effectively
treating brain tumors hinges on early and precise dis-
ease diagnosis. Decisions regarding treatment methods
are influenced by factors such as the tumor’s pathologi-
cal type, grade, and stage at diagnosis. Neuro-oncologists
have harnessed computer-aided diagnostic (CAD) tools
for various purposes, including tumor detection, catego-
rization, and grading within the realm of neurology [7].
A glioma is a type of tumor that originates in brain
tissue, distinct from nerve cells or blood vessels. In
contrast, meningiomas develop from the protective
membranes that envelop the brain and central nervous
system, while pituitary tumors grow within the confines
of the skull. Among these three tumor types, menin-
giomas are relatively rare and generally benign. Con-
versely, gliomas constitute the most prevalent form of
malignant brain tumors. Even though pituitary tumors
may be benign, they can still give rise to significant
medical complications [8]. Brain tumors rank as a lead-
ing cause of mortality worldwide. Research underscores
the significance of early and accurate identification,
coupled with prompt treatment, in improving sur-
vival rates for patients with cancerous tumors. In cer-
tain instances, healthcare professionals may encounter
the need to differentiate between strokes and tumors.
Hence, the early detection of brain tumors assumes piv-
otal importance for providing effective care and poten-
tially extending the affected individual’s lifespan [9].
Convolutional neural networks (CNNs), distinguished
by their multi-layered architecture and high diagnos-
tic accuracy when provided with ample input images,
currently stand as a highly effective approach in image
processing. Neural networks, including auto-encoders,
an unsupervised learning technique, are harnessed for
representation learning [10]. Magnetic resonance imag-
ing (MRI) emerges as an exceptional tool for obtain-
ing clear and detailed visualizations within the human
body. Unlike X-rays or CT scans that involve ionizing
radiation, MRI offers significantly enhanced contrast
between various soft tissues. Moreover, MRI technol-
ogy furnishes detailed images from multiple angles,
providing radiologists with abundant data on human
soft-tissue anatomy [11]. The aim of this paper is to
introduce three fully automatic CNN models designed
for the multi-classification of brain tumors, utilizing
publicly available datasets. To the best of the authors’
knowledge, this represents the first endeavor in multi-
classifying brain tumors from MRI images using CNNs,
wherein nearly all the hyperparameters are automati-
cally tuned through the grid search optimizer. The rest
of this paper is organized as follows: Introduction Sec-
tion: this section provides a comprehensive overview
of various tumor types and their diagnostic methods;
Related work Section: in this section, we delve into
recent articles, examining their methods, outcomes,
and applications; Materials and methods Section: here,
we detail the utilization of datasets and describe the
proposed model architectures; Experimental study
Section: this section centers on a comparative analy-
sis of the accuracies achieved by our proposed method
and other state-of-the-art approaches; Conclusions

of 21
Section: this section offers the concluding remarks and
insights related to our proposed model.
Related work
The author’s goal was to devise a classification approach
that is notably more accurate, cost-effective, and self-
training, utilizing an extensive collection of authentic
datasets rather than augmented data. The customized
VGG-16 (Visual Geometry Group) architecture was
employed to classify 10,153 MRI images into three dis-
tinct classes (glioma, meningioma, and pituitary). The
network demonstrated a remarkable performance,
achieving an overall accuracy of 99.5% and precision
rates of 99.4% for gliomas, 96.7% for meningiomas, and
100% for pituitaries [12]. The proposed model’s efficacy
was assessed using three CNN models: AlexNet, Visual
Geometry Group (VGG)-16, and VGG-19. AlexNet
achieved a peak detection accuracy of 99.55% using 349
images sourced from the Reference Image Database to
Evaluate Response (RIDER) neuro MRI database. For
brain tumor localization, employing 804 3D MRIs from
the Brain Tumor Segmentation (BraTS) 2013 database,
a Dice score of 0.87 was achieved [13]. In the investiga-
tion of brain tumor categorization, an array of deep- and
machine-learning techniques, including softmax, Ran-
dom Forest, Support Vector Machine (SVM), K-Nearest
Neighbors, and the ensemble method, were employed.
These outcomes were compared with existing methods.
Notably, the Inception-v3 model exhibited the highest
performance, attaining a test accuracy of 94.34%. This
advancement holds the potential to establish a promi-
nent role in clinical applications for brain tumor analysis
[14]. An effective approach was proposed for categoriz-
ing brain MRIs into four classes: normal and three forms
of malignant brain tumors (glioblastoma, sarcoma, and
metastatic bronchogenic carcinoma). The method inte-
grates the discrete wavelet transform (DWT) with a
deep neural network (DNN). Employing a deep neural
network classifier, one of the DL designs, a dataset of 66
brain MRIs was classified into the specified categories.
The integration of DWT, a powerful feature extraction
technique, principal component analysis (PCA), and the
classifier yielded commendable performances across all
evaluation metrics [15]. The author introduced a strategy
involving a CNN to distinguish brain tumors from 2D
MRI scans of the brain. This initial separation is subse-
quently followed by the application of conventional clas-
sifiers and DL techniques. In addition, an SVM classifier,
along with various activation algorithms, such as soft-
max, RMSProp, and sigmoid, were employed to validate
and cross-check the proposed approach. The implemen-
tation of the author’s suggested solution was executed
using TensorFlow and Keras in the Python programming
language, chosen for its robust capabilities in expediting
tasks. The achieved accuracy rate for the CNN model
stood at an impressive 99.74% [16]. This paper presents
a brain tumor classification approach employing open-
access datasets and CNN techniques. The methodology
utilizes open-access datasets to classify tissue as either
tumor or non-tumor through a distinctive framework that
combines discrete cosine transform-based image fusion,
CNN super-resolution, and a classifier. Employing super-
resolution and the ResNet50 architecture, the framework
attained an impressive accuracy of 98.14% [17].
A novel approach for dimensionality reduction is pro-
posed, utilizing the Grey Wolf Optimizer (GWO) and
rough-set theory. This method identifies relevant features
from extracted images, distinguishing between high-
grade (HG) and low-grade (LG) glioblastoma multiforme
(GBM) while accommodating feature correlation con-
straints to eliminate redundant attributes. Additionally,
the article introduces a dynamic architecture for mul-
tilevel layer modeling in a Faster R-CNN (MLL-CNN)
approach. This is achieved using a feature weight factor
and a relative description model to construct selected fea-
tures, thereby streamlining the processing and classifying
of long-tailed files. This advancement leads to improved
training accuracies for CNNs. The findings illustrate that
the overall survival prediction for GBM brain growth
achieves a higher accuracy of 95% and a lower error rate
of 2.3% [18]. The work involves the classification of 253
high-resolution brain MR images into normal and path-
ological classes. To efficiently and accurately train deep
neural models, MR images were scaled, cropped, pre-
processed, and enhanced. The Lu-Net model is compared
against LeNet and VGG-16 using five statistical met-
rics: precision, recall, specificity, F-score, and accuracy.
The CNN models were trained on enhanced images and
validated on 50 sets of untrained data. LeNet, VGG-16,
and the proposed approach achieved accuracy rates of 88%,
90%, and 98%, respectively [19]. MIDNet18 outperformed
AlexNet in categorizing brain tumor medical images. The
proposed MIDNet18 model demonstrated effective learn-
ing, achieving a binary classification accuracy exceeding
98%, which is statistically significant (independent-sample
t-test, p < 0.05). MIDNet18 excelled across all the perfor-
mance indicators for the dataset used in this study [20].
The objective of this study was to facilitate accurate
early-stage diagnoses by medical professionals. Three DL
architectures—AlexNet, GoogLeNet, and ResNet50—
were employed to identify brain tumor images. Among
them, the ResNet50 architecture demonstrated the
highest accuracy rates. The experimental results yielded
an accuracy of 85.71%, with the potential for further

of 21
enhancement in future research [21]. In the realm of
Alzheimer’s disease diagnosis, the CNN approach was
utilized to detect patients using MRSI and supplemen-
tary MRI data. High Matthews Correlation Coefficient
(MCC) scores were achieved, with area-under-the-curve
values of 0.87 and 0.91 for MRSI and MRI, respectively.
A comparative analysis highlighted the superiority of
Partial Least Squares and Support Vector Machines. The
proposed system automatically selected critical spectral
regions for diagnosis, corroborating findings with lit-
erature biomarkers [22]. CNNs, ML pipelines inspired
by biological neural processes, have been extensively
studied. The author’s approach involved first acquir-
ing an understanding of CNNs, followed by a literature
search for a segmentation pipeline applicable to brain
tumor segmentation. Additionally, the potential future
role of CNNs in radiology was explored. The applica-
tion of CNNs was demonstrated in predicting survival
and medication responses through analyses of the brain
tumor shape, texture, and signal intensity [23]. In this
paper, the state-of-the-art object detection framework
YOLO (You Only Look Once) was employed to identify
and classify brain tumors using DL. YOLOv5, a revo-
lutionary object detection algorithm, stood out for its
computational efficiency. The RSNA-MICCAI brain
tumor radiogenomics classification BraTS 2021 dataset
served as the basis. YOLOv5 achieved an 88% precision
rate [24]. The primary aim of this method is to classify
brain images as healthy or tumorous using test MRI data.
MRI-based brain tumor research offers superior internal
imaging compared to CT scans. The approach involves
denoising MRI images with an anisotropic diffusion fil-
ter, segmenting using morphological operations, and
classifying via a five-layer CNN-based hybrid technique,
outperforming other methods. The developed model,
utilizing the publicly available KAGGLE brain MRI data-
base, achieved an accuracy rate of 88.1% [25]. The adop-
tion of AI-powered computer systems can assist doctors
in making more accurate diagnoses. In this research,
we developed a brain tumor diagnostic system based
on CNN technology, utilizing Ranger optimization and
the extensive pre-processing of data from the Efficient-
Netv2 architecture [26]. This research introduces a novel
topology for a parallel deep CNN (PDCNN) designed to
extract both global and local features from two parallel
stages. Overfitting is addressed through the utilization of
dropout regularization and batch normalization. Unlike
conventional CNNs that collect features randomly with-
out considering local and global contexts, our proposed
PDCNN architecture aims to capture a comprehensive
range of features [27]. This study focuses on the classi-
fication of meningiomas, gliomas, and pituitary tumors
using MRI imaging. The Dual VGG-16 CNN, equipped
with a proprietary CNN architecture, constitutes the
DCTN mode [28]. The importance of the early detection
of brain tumors cannot be overstated. Biopsies of brain
tumors, the gold standard for diagnosis, are only possi-
ble during life-altering brain surgery. Methods based on
computational intelligence can aid in the diagnosis and
categorization of brain tumors [29]. The author employed
a DL model to classify MRI scans into glioma and normal
categories, preceded by the extraction of scan informa-
tion. Convolutional recurrent neural networks (CRNNs)
were utilized for generating the classifications. This sug-
gested method significantly improved the categorization
of brain images within a specified input dataset [30]. The
network was trained and tested using BraTS2019 data.
The approach was evaluated using the Dice similarity
coefficient (DSC), sensitivity (Sen), specificity (Spec), and
Hausdorff distance (HD). The DSCs for the entire tumor,
tumor core, and enhancing tumor were 0.934, 0.911, and
0.851, respectively. The subregion Sen values were 0.922,
0.911, and 0.867. The Spec and HD scores were 1.000,
1.000, and 3.224, 2.990, 2.844, respectively [31]. The can-
cer region segmentation from brain images is achieved
using Deep K-Net, a hybrid approach that combines
K-Net and utilizes Deep Joint Segmentation with Ruzicka
similarity. The K-Net is trained using a Driving Training
Taylor (DTT) algorithm. The DTT algorithm optimizes
the Shepard CNN (ShCNN) for classification [32].
The author provided an overview of contemporary
computer-aided detection methods that utilize WCE
images as input, distinguishing them as either diseased/
abnormal or disease-free/normal. We conducted an
evaluation of approaches designed for the detection of
tumors, polyps, and ulcers, as these three conditions
are categorized similarly. Furthermore, because general
abnormalities and bleeding within the GI tract could be
indicative of these disorders, we made an effort to shed
light on the research conducted for the identification of
abnormalities and bleeding within WCE images [33].
Author have included several research studies, each
accompanied by detailed descriptions of their techniques,
findings, and conclusions. Additionally, we provide a
discussion and comparison of previous review articles,
which serves as a reference point for the current survey,
while also highlighting its limitations [34]. To enhance
feature extraction, our proposed deep CNN model intro-
duces an innovative approach by incorporating multiple
convolutional kernels with varying window widths within
the same hidden layer. This architecture is designed to be
lightweight, consisting of 16 convolutional layers, 2 fully
connected layers (FCN), and a softmax layer serving as
the output layer. The activation function employed in the
first 15 layers is MISH, followed by the Rectified Linear
Unit (ReLU) activation function. This combination not

of 21
only facilitates profound information propagation but
also offers self-regularized, smoothly non-monotonic
characteristics, while effectively mitigating saturation
issues during training. The authors present a comprehen-
sive set of experimental results, comparing our model’s
performance against benchmarks like the MICCAI 2015
challenge and other publicly available datasets. Our
findings demonstrate that the proposed model excels in
terms of accuracy, sensitivity, the F1-score, the F2-score,
and the Dice coefficient [35].
Materials and methods
Materials
The study used four different datasets that can be found
in freely accessible databases. The Figshare dataset is the
name of the first dataset. From 19 patients with glioblas-
tomas (G-IV), MRI multi-sequence images were taken
and added to the Figshare dataset, which is a targeted
collection of data. There are a total of 70,221 images con-
tained within this collection. The name of the second
collection of data is the Repository of Molecular Brain
Neoplasia Data (REMBRANDT) [36]. This set of data has
MRI images of gliomas with grades II, III, and IV from
133 patients, and it has 109,021 images in total.
The Cancer Genome Atlas Low-Grade Glioma data-
set is the third dataset that was analyzed (TCGA-LGG)
[37], and it has 242,185 MRI images of patients with
low-grade gliomas (G-I and G-II) and incorporates data
from 198 patients. These three datasets are part of the
Cancer Imaging Archive (TCIA) project [38]. In each
instance, multimodal imaging was performed, including
T1-contrast-enhanced and FLAIR images [39]. The last
collection of data used in this investigation consists of
3067 T1-weighted, contrast-improved images from 243
patients with three different types of brain tumors: glio-
mas (1427 slices), meningiomas (709 slices), and pituitary
tumors (931 slices). Figure 1 depicts the different grades
of brain tumors from the dataset. Totally, 3165 images
are collected for the Classification-1 mode, 1743 of which
are malignant tumors and 1422 of which are not. For the
Classification-2 mode, 4195 images are collected. There
are 910 normal images, 985 glioma images, 750 menin-
gioma images, 750 pituitary images, and 800 metastatic
images. For the Classification-3 mode, we obtain a total
of 4720 images: 1712 G-II, 1296 G-III, and 1712 G-IV.
Table 1 represents the dataset split-up details for the pro-
posed model.
Methods
Convolutional neural network
The CNN is the neural network DL model that is most
frequently employed. A common CNN model has two
components: classification and feature extraction. A
CNN architecture has five key layers: the input layer,
convolution layer, pooling layer, fully connected layer,
and classification layer. The CNN provides the extraction
and classification of features using successively arranged
trainable layers. Convolutional and pooling layers are
typically included in the feature extraction phase of a
CNN, whereas fully connected and classification layers
are typically included in the classification part. This pro-
posed study suggests creating three fully automatic CNN
models for classifying different types of brain tumors
using MRI images. Grid search optimization tunes the
key hyperparameters of the CNN models automatically.
The primary of these CNN models determines whether a
particular MRI image of a patient has a tumor or not, as
it is employed to diagnose brain tumors. Throughout this
study, this mode will be referred to as “Classification 1”
(C-1). According to Fig. 2, the proposed CNN model for
C-1 consists of thirteen weighted layers: one input layer,
two convolution layers, two ReLU layers, one normaliza-
tion layer, two max-pooling layers, two fully connected
layers, one dropout layer, one softmax layer, and one clas-
sification layer.
The initial CNN model is meant to classify an image
into two groups, and it has two neurons in the output
layer. Finally, a softmax classifier is fed the output of the
fully connected layer (a two-dimensional feature vector)
to determine whether a tumor is present or not. Table 2
illustrates detailed information on the CNN model.
There are five distinct forms of brain tumors that are dis-
tinguished by the second CNN model: benign, malignant,
meningioma, pituitary, and metastatic. Throughout this
study, this mode will be referred to as “Classification 2”
(C-2). As shown in Fig. 3, the proposed CNN model for
C-2 contains a total of 25 weighted layers: 1 input layer, 6
convolution layers, 6 ReLU layers, 1 normalization layer,
6 max-pooling layers, 2 fully connected layers, 1 drop-
out layer, 1 softmax layer, and 1 classification layer. The
output layer of the second CNN model has five neurons
as a result of the model’s intention to classify each given
image into five distinct categories. The final prediction of
the tumor type is made using a softmax classifier, which
receives as input the five-dimensional feature vector gen-
erated by the final fully connected layer. Table 3 illus-
trates detailed information on the CNN model. The third
proposed CNN framework divides glioma brain tumors
into three grades, which are called G-II, G-III, and G-IV.
Throughout this study, this mode will be referred to as
“Classification 3” (C-3). As can be seen in Fig. 4, the pro-
posed CNN model for C-3 consists of a total of sixteen
weighted layers: one input layer, three convolution layers,
three ReLU layers, one normalization layer, three max-
pooling layers, two fully connected layers, one dropout
layer, one softmax layer, and one classification layer. The

of 21
most recent CNN model has three neurons in the out-
put layer because it is meant to divide every image into
three groups. The final fully connected layer, which is a
three-dimensional feature vector, is sent to the softmax
classifier as an input. The softmax classifier then makes a
final prediction about the tumor grade. Table 4 illustrates
detailed information on the CNN model.
Performance metric evaluation
It is essential to analyze the classification performance in
image classification research to provide a rational founda-
tion for the outcomes of the investigation. Many different
performance evaluation metrics have been used for an
extended period in studies involving image classification
and that have evolved into standard performance evalu-
ation metrics in studies that are similar to the prior. The
proposed model used different parametric methods for
evaluation, such as precision, sensitivity, and accuracy.
These measures, which are generally acknowledged as
standard performance evaluation metrics in image classifi-
cation research, are also employed in this article in order
to measure the accuracy and reliability of the classification
process. Furthermore, the receiver operation characteris-
tic (ROC) curve area, also known as the AUC of the ROC
Fig. 1 a Manual tumor segmentation; b WHO grade II (first row), grade III (second row), and grade IV (third row) brain tumors

of 21
curve, is used to evaluate the models’ performance. The fol-
lowing are the equations containing the corresponding for-
mulas for each of these measurements:
(1)
Accuracy =
∅ + β
∅ + β + α + γ
(2)
Specificity =
β
β + α
(3)
Precision =
∅
∅ + α
where ø is true positive, β is true negative, α is false posi-
tive, and γ is false negative.
Experimental Study
We implemented the proposed classification model in
MATLAB2021a on a computer with the specifications of
32 GB RAM and an Intel E3-1245v6 @3.70GHz CPU.
Optimization of the Hyperparameters
There have been several developments in the field of med-
ical image processing that have led to the increased use
of CNNs, and, as a result, some challenges have arisen in
their use. The designs designed to obtain more effective
outcomes are deeper, and the input images are becoming
higher-quality, which leads to an increase in the amount
of processing resources required. Sufficient hardware and
tuning the network’s hyperparameters are essential for
lowering these computing costs and maximizing results.
As a result, the proposed CNN models have nearly all of
their essential hyperparameters automatically set using
the grid search optimization technique. When the search
space for possible values is small, grid search optimiza-
tion is a great way to improve a CNN’s hyperparameter
optimizations. The grid search can select the superior
one by training the network through a wide range of
possible combinations. CNN models have architectures
that are quite complicated and that have a lot of hyper-
parameters. In most cases, these hyperparameters can
(4)
Sensitivity =
∅
∅ + γ
Table 1 Number of MRI images in the dataset
Dataset Split-Up
Classification No. of Images in the
Group
Total
No. of
Images
Mode Group
I Malignant 1743 3165
Non-malignant 1422
II Benign 910 4195
Glioma 985
Meningioma 750
Pituitary 750
Metastatic 800
III G-II 1712 4720
G-III 1296
G-IV 1712
Fig. 2 Proposed CNN model architecture for“C-1”mode

of 21
be arranged into two distinct categories: architectural
hyperparameters and fine-adjustment hyperparameters.
Architectural hyperparameters include the following:
the number of convolutional pooling layers, the number
of fully connected layers, the number of filters, the filter
sizes, and the activation function. The regularization,
momentum, minibatch size, and learning rate are among
the fine-adjustment hyperparameters. In the current
analysis, the hyperparameters of the architecture are ini-
tially tuned using Algorithm 1.
Table 2 Detailed information on CNN model employed for“C-1”mode
Layer Name CNN Layer Activations Parameters (Trainable) Total No. of
Trainable
Parameters
Input 227×227×3 227×227×3 nil 0
Convolutional 128 (6×6×3), stride of (4,4), with (0 0 0 0) padding 56×56×128 6×(6×3)×128 weights, 1×1×128 bias 13,954
Activation layer Activation layer-1 56×56×128 nil 0
Normalization Normalization (cross-channel) 56×56×128 nil 0
Max_pooling (2×2) with stride of (2,2), and (0 0 0 0) padding 28×28×128 nil 0
Convolutional 96 (6×6×128), stride of (1,1), and (2 2 2 2) padding 31×31×96 2×(2×128)×96 weights, 1×1×96 bias 49,246
Fully_connected 512 Fully_connected 1×1×512 512×21,700 weights, 512×1 bias 11,060,714
Dropout 30% 1×1×512 nil 0
Fully_connected 2 Fully_connected 1×1×2 512×2 weights, 2×1 bias 1026
Softmax Softmax 1×1×2 nil 0
Classification Tumor or non-tumor nil nil 0

of 21
Algorithm 1. Architectural hyperparameters will be optimized using
a grid search algorithm
After determining the architectural hyperparameters,
Algorithm 2 is used to optimize the fine-adjustment hyper-
parameters. In this proposed study, the grid search is
carried out on the training set employing a fivefold cross-
validation method. The dataset is split into five different
sets. Four of these sets are used for training, and the fifth
set is used for testing. For the Classification-1 mode, there
are 3165 images, for the Classification-2 mode, there are
4195 images, and for the Classification-3 mode, there are
4720 images. For each classification mode, the dataset is
randomly split into a training set, a validation set, and a test
set, with the ratio being 60:20:20. Basically, the grid search
method goes through each possible setting for each param-
eter and finds the one that gives the best performance. In
order to obtain the highest possible degree of accuracy
with Algorithm 1, there are five parameters that need to be
improved.
Trainable
Parameters
Input 227×227×3 227×227×3 nil 0
Fully_connected 512 Fully_connected 1×1×512 512×32 weights, 512×1 bias 16,896
Classification Benign, glioma, pituitary, metastatic, and meningioma nil nil 0

of 21
Algorithm 2. Architectural hyperparameters will be optimized using
a grid search algorithm
Many possible combinations for these parameters,
including 4, 4, 7, 5, and 4, correspondingly. As a result,
the total number of possible permutations to be exam-
ined is 4 × 4 × 7 × 5 × 4, which equals 2240. Because
2240 combinations need to be checked using the fivefold
cross-validation technique, the grid search algorithm cre-
ated to optimize the CNN model hyper-parameters is
carried out 11,200 times. Similar to the first algorithm,
the second algorithm has four parameters that need to be
optimized to achieve the highest level of accuracy. A wide
range of permutations are possible in these parameters,
for example, 4, 4, 5, and 4. As a result, the total number of
possible permutations that need to be examined is 4 × 4
× 5 × 4, which equals 320. Because 320 different possible
combinations need possible combinations that need to
be tested using the fivefold cross-validation method, the
grid search technique developed to improve the correc-
tion hyperparameters of the CNN model is carried out a
total of 1600 times. As shown in Tables 5, 6 and 7, the
grid search optimization algorithm found the best possi-
ble values for the hyperparameters of the C-1, C-2, and
C-3 modes.
Optimized Convolutional Neural Network Outcomes
The fivefold cross-validation approach for the C-1
mode is utilized to conduct the proposed model’s per-
formance analysis. The dataset is partitioned into five
different sets, four of which are utilized for training
purposes, while the fifth set is placed to use for testing
purposes. There are five total iterations of the experi-
ments, and the classification performance of the mode
is evaluated for each fold, and then the overall model’s
average classification performance is computed. High
accuracy results from the training and validation phases
are meaningless if the trained and hyperparameter-
tuned CNN is not tested on its ability to predict sam-
ples that have not yet been seen. Hence, to assess the
effectiveness of the trained CNN to assess the trained

of 21
CNN’s effectiveness on predicting samples, a test data-
set is randomly allocated and segregated alongside the
training and validation datasets. If this step is skipped,
the high accuracy may result from biased dataset
assignment. Table 8 displays the results of randomly
splitting the 3165 images from the study into the train-
ing, validation, and test sets in the ratio of 60:20:20 for
the C-1 mode.
A total of 299 images are taken randomly from the
dataset for each category, and then those images are used
for testing. The activations of the CNN’s convolution
layers can be displayed for a better view of the features
that the CNN has learned due to its training. With this
representation, the researcher may easily observe the
network’s progress. Figures 5 and 6 each depict the acti-
vations of the first and second convolutional layers. One
of the images in the grid serves as a representation of the
channel’s outcome. White areas represent highly posi-
tive activations, while grey areas represent moderately
activated channels. While the first convolutional layer
of the CNN is used to learn features such as color and
edges, the second convolutional layer is used to learn
more complex information, such as the borders of brain
tumors. The succeeding (deeper) convolutional layers
Trainable
Parameters
Input 227×227×3 227×227×3 nil 0
Fully_connected 512 Fully_connected 1×1×512 512×6144 weights, 512×1 bias 3,146,240
Classification G-II, G-III, G-IV nil nil 0
Table 5 The grid search-yielded optimal results for the
hyperparameters for the C-1 mode
Hyperparameters Changes in Parameter Values Maximal Value
Layers of maximum
pooling and CNN
(1, 2, 3, 4) 2
Number of layers
that are completely
connected
(1, 2, 3, 4) 2
Total number of filters (8, 16, 24, 32, 48, 64, 96, 128,
256)
64, 96, 128
Intensity of filtration (3, 4, 5, 6, 7) 6, 6
Role of activation (ReLU, ELU, Leaky ReLU) ReLU
Size of minibatch (4, 6, 16, 24, 32, 64) 32
Rate of change (0.78, 0.77, 0.95, 0.96) 0.95
Rate of learning (0.0002, 0.00043, 0.002, 0.004) 0.0002
R2—regularization (0.0002, 0.00043, 0.002, 0.004) 0.0002
Table 7 The grid search-yielded optimal results for the
hyperparameters for the C-3 mode
Layers of maximum
pooling and CNN
(1, 2, 3, 4) 3
Number of layers
that are completely
connected
(1, 2, 3, 4) 2
Total number of filters (8, 16, 24, 32, 48, 64, 96, 128,
256)
64, 96, 128
Intensity of filtration (3, 4, 5, 6, 7) 6, 6, 4
Rate of change (0.78, 0.77, 0.95, 0.96) 0.95
Rate of learning (0.0002, 0.00043, 0.002, 0.004) 0.004

of 21
build up their features by merging the features learned by
the earlier convolutional layers.
Figure 5 shows 96 of the 128 channels in the CNN’s
first convolutional layer running in C-1 mode. This layer
contains a total of 128 channels. Figure 6 shows an image
of the second convolutional layer of the network, which
has 96 channels. Every layer of the CNN is composed of
channels, which are arrays in two dimensions. One of
the images in Fig. 5 represents the output of each chan-
nel in the first convolutional layer. In these images, strong
positive activations are shown by white pixels, and strong
negative activations are shown by black pixels. Similarly,
grey pixels on the input image indicate channels that are
not highly active. Figure 7 depicts the activations of a
particular channel and the channel with the most signifi-
cant activation in the first convolutional layer. The pres-
ence of white pixels in the channel of Fig. 7 demonstrates
that this channel is highly activated at the tumor’s loca-
tion. Although the CNN was never instructed to learn
about tumors, it is possible to conclude that it has picked
up on the fact that tumors have distinguishing qualities
that allow it to differentiate between different categories
of images.
These CNNs are able to discover helpful character-
istics on their own, unlike earlier artificial neural net-
work methods that typically required manual design to
fit a particular mode. In this proposed article, learning
to recognize tumors improves the ability to distinguish
between a tumor image and non-tumor image. After
the process of classification has been completed, the
efficiency of the CNN models must be evaluated using
different reliable approaches. The metrics, like the speci-
ficity, sensitivity, precision, and accuracy measures, as
well as the area under the ROC curve, are used to per-
form the performance evaluation of the proposed model.
The proposed CNN’s loss and accuracy plots for the C-1
mode are shown in Fig. 8. After 340 iterations, the model
proposed for C-1 was able to classify with a 99.53% accu-
racy. It is pretty clear, as shown in Fig. 8, that approxi-
mately 250 iterations are required to reach an almost
perfect level of accuracy. Figure 9 depicts the confusion
matrix for the Classification-1 mode. As can be seen
in Fig. 10, the area under the ROC curve has a value of
0.9995 for its AUC. The results presented here demon-
strate that the recommended CNN model is capable of
identifying brain tumors. Table 9 shows the measures of
Table 6 The grid search-yielded optimal results for the hyperparameters for the C-2 mode
Layers of maximum pooling and CNN (1, 2, 3, 4) 6
Number of layers that are completely connected (1, 2, 3, 4) 2
Total number of filters (8, 16, 24, 32, 48, 64, 96, 128, 256) 16, 24, 32, 48, 64, 96, 128
Intensity of filtration (3, 4, 5, 6, 7) 6, 6, 4, 6, 2, 6
Rate of change (0.78, 0.77, 0.95, 0.96) 0.95
Rate of learning (0.0002, 0.00043, 0.002, 0.004) 0.0002
Table 8 Training, validating, and testing phases of proposed CNN model
Dataset Split-Up Training, Validation, and Testing Modes
Classification No. of Images in the
Group
Total No. of Images Training Mode
(60%)
Validation Mode
(20%)
Test
Mode
(20%)
Task Group
I Malignant 1743 3165 1899 633 633
Non-malignant 1422
II Benign 910 4195 2517 839 839
Glioma 985
Meningioma 750
Pituitary 750
Metastatic 800
III G-II 1712 4720 2832 944 944
G-III 1296
G-IV 1712

of 21
the accuracy, such as the true positive (TP), true negative
(TN), false positive (FP), false negative (FN), accuracy
(Acc), specificity (Sp), sensitivity (Se), and precision (Pr).
Figure 10 depicts the ROC curve for the Classification-1
(C-1) task.
Figure 11 shows the results of the classification and
the predicted probabilities for each of the four tests
conducted in C-1 mode. Implementing the fivefold
cross-validation method for the C-2 mode evaluates
the effectiveness of the proposed framework. The data-
set is partitioned into five sets, four of which are utilized
for training purposes, while the fifth set is placed for
testing purposes. There are five total iterations of the
experiments. The classification performance of the job is
evaluated for each fold, and then the overall model’s aver-
age classification performance is computed. As indicated
in Table 8, there are sufficient images for the C-2-mode
training, validation, and test sets to be randomly divided
in a ratio of 60:20:20 for a sample size of 4195. From the
dataset of each class that will be used to test the model,
158 images are randomly selected to be removed. The
accuracy and loss plots of the suggested CNN model for
Fig. 5 First CNN activation layer for C-1 mode
Fig. 6 Second CNN activation layer for C-1 mode

of 21
the C-2 task are displayed in Fig. 12. The proposed CNN
method for the C-2 mode achieves a 93.81% accuracy in
classification after 294 iterations. As shown in Fig. 13, the
area under the ROC curve has a value of 0.9981. These
findings demonstrate the proposed CNN model’s capa-
bility to classify brain tumor types. Figure 14 depicts a
confusion matrix, and Table 9 lists the many measures
of precision, such as TP, TN, FP, FN, Acc, Sp, Se, and Pr.
According to Table 9, an accuracy of 97.26% is attained
when classifying a glioma, 97.50% when classifying a
meningioma, 96.86% when classifying metastasis, 97.99%
when classifying a healthy brain, and 95.59% when clas-
sifying the pituitary tumor type for the C-2 mode.
Figure 14 depicts the ROC curve for the Classification-2
(C-2) task.
The fivefold cross-validation process for the C-3 mode
is utilized to evaluate the efficacy of the proposed models.
The dataset is partitioned into five different sets, out of
which four are used for training and the fifth is used for
testing. There are five total iterations of the experiments.
Following an analysis of the classification performance
of the mode for each fold, an average classification per-
formance for the model is computed. For the C-3 mode,
sufficient images can be randomly divided into training,
validation, and test sets in the proportions 60:20:20, as
indicated in Table 8, randomly excluding three hundred
Fig. 7 C-1-mode strongest and moderate images from original input image
Fig. 8 C-1-mode accuracy and loss curves

of 21
and five images from the dataset of each class to be uti-
lized to evaluate the model. The loss and accuracy graphs
of the proposed CNN for the C-3 mode are shown in
Fig. 15. Figure 16 depicts the confusion matrix for the
C-3 mode. The proposed approach for the C-3 mode
obtains a classification accuracy of 98.16% after 344 iter-
ations. Figure 17 depicts the ROC curve for the Classi-
fication-3 (C-3) task. Table 9 shows that an accuracy of
98.16% is reached when classifying grade II, 100% when
classifying grade III, and 98.17% when classifying grade
IV for brain tumor grades in the C-3 mode. The three
different classification outcomes of the proposed CNN
model were compared with other conventional CNN
approach outcomes to evaluate the proposed system clas-
sification ability. To achieve this goal, the same experi-
ments were performed with the same dataset, utilizing
well-known and popular pretrained CNN models, such
as AlexNet, DenseNet121, ResNet-101, VGG-16, and
Fig. 9 C-1 confusion matrix
Fig. 10 C-1-mode average of ROC curve

of 21
Table 9 Proposed CNN model parameter metric outcomes for all classification modes
Metrics Classes TP TN FP FN Acc Sp Se Pr Total
I Malignant 268 326 3 0 99.50 99.09 100.00 98.89 268
Non-malignant 326 268 0 3 99.50 100.00 99.09 100.00 329
II Benign 183 598 8 8 97.99 98.68 95.81 95.81 191
Glioma 132 650 10 12 97.26 98.48 91.67 92.96 144
Meningioma 138 643 6 14 97.50 99.08 90.79 95.83 152
Pituitary 160 598 24 11 95.59 96.14 93.57 86.96 171
Metastatic 127 643 11 14 96.86 98.32 90.07 92.03 141
III G-II 332 574 9 8 98.16 98.46 97.65 97.36 340
G-III 248 679 0 0 100.00 100.00 100.00 100.00 248
G-IV 330 580 8 9 98.17 98.64 97.35 97.63 339
Fig. 11 The results of classification and predictions for the probabilities of four different test images for the C-1 mode

of 21
GoogleNet. Table 10 illustrates the performance metric
outcome comparison of the proposed CNN model with
existing CNN approaches. Figure 18 depicts the graphi-
cal representation of the proposed and existing models’
result comparison.
The results shown in Table 10 illustrate that the
proposed CNN models outperform other networks in
every classification mode. The pretrained DenseNet121
model, which obtains a 93.89% classification accuracy
in the brain tumor detection test (C-1 mode), is the
model that is closest to the suggested model. The pre-
trained VGG-16 model obtains an 89.19% accuracy in
the brain tumor-type classification mode (C-2 mode).
It is the model that is closest to the proposed CNN
model. After the proposed CNN model, the pretrained
GoogleNet model achieves a classification accuracy of

of 21
95.12%, making it the best network available for grad-
ing tumors (C-3 mode). It is clear that the proposed
CNN models are better than the pretrained networks,
which were built and trained using generic datasets and
methods for a wide range of image classification tasks.
Table 11 illustrates the proposed and existing model
outcome comparison. The proposed CNN models, con-
versely, were designed to deal with more specific issues,
like identifying and defining various types and stages of
brain tumors. Finally, MRI images of brain tumors are
used to train and evaluate the proposed models.
Conclusions
In this research, we propose a multi-classification
method for identifying brain tumors at an early stage
using (CNN) models, in which nearly all the hyperpa-
rameters are automatically optimized via grid search.
By using publicly available medical imaging datasets,
three reliable CNN models have been designated
to perform three distinct brain tumor classification
tasks. A high level of accuracy, such as 99.53%, can be
attained in the process of detecting brain tumors. In
addition, a remarkable accuracy of 93.81% is achieved

of 21
Table 10 Performance metric outcome comparison of the proposed CNN model with existing CNN approaches
CNN Models C-1 Mode C-2 Mode C-3 Mode
Acc (%) AUC Acc (%) AUC Acc (%) AUC
GoogleNet 74.21 0.8108 77.89 0.8212 95.12 0.9617
AlexNet 89.23 0.8989 84.24 0.8501 91.08 0.9772
DenseNet121 93.89 0.9412 77.67 0.8122 86.07 0.8809
ResNet101 93.29 0.9442 76.45 0.8115 86.42 0.881
VGG-16 88.87 0.9201 89.19 0.8112 84.87 0.8663
Proposed CNN approach 99.53 0.9994 93.81 0.9984 98.56 0.9993
Fig. 18 Graphical illustration of proposed and existing models’outcome comparison

of 21
when classifying brain MR images into the catego-
ries of glioma, meningioma, pituitary, normal brain,
and metastatic. The final step is grading glioma brain
tumors, which can be performed with an accuracy of
98.56% for grades II, III, and IV. A good number of
medical images are used to train and test the CNN
models that are being proposed. Results from the
proposed CNN models and comparisons with cur-
rent methods show that CNN models made with the
proposed optimization framework work well. In this
work, CNN models were made that can help clinicians
and radiologists check primary screenings for multi-
ple types of brain tumors.
Acknowledgements
Not applicable.
Institutional Review Board Statement
Not applicable.
Human Participants Research Checklist:
Complete the following if your study involved human participants or human
participants’data. These questions should be addressed for prospective and
retrospective studies.
1. Did you obtain ethics approval for this study?
• If yes, please upload (file type“Other”) the original approval document you
received from your ethics committee. If the original document is in another
language, please also provide an English translation.
Response: N/A
• If you did not obtain ethical approval, please explain why this was not
required below.
2. If you prospectively recruited human participants for the study for example,
you conducted a clinical trial, distributed questionnaires, or obtained tissues,
data or samples for the purposes of this study, please report in the Methods:
i. the day, month and year of the start and end of the recruitment period for
this study.
ii. whether participants provided informed consent, and if so, what type was
obtained (for instance, written or verbal, and if verbal, how it was documented
and witnessed). If your study included minors, state whether you obtained
consent from parents or guardians. If the need for consent was waived by the
ethics committee, please include this information.
Response: N/A
3. If you are reporting a retrospective study of medical records or archived
samples, please report in the Methods section:
i. the day, month and year when the data were accessed for research purposes
ii. whether authors had access to information that could identify individual
participants during or after data collection
Response: N/A
Authors’contributions
Conceptualization, S.S. and D.F.; methodology, S.K.M. and H.R.; valida-tion, H.R.
and M.A.S.; data curation, B.D.S.; writing—original draft, S.S. and D.F.; writing—
review and editing, S.K.M. and H.R.; visualization, B.D.S; supervision S.K.M., and
M.A.S.; project ad-ministration, S.K.M., and M.A.S. All authors have read and
agreed to the published version of the manuscript.
Funding
This research received no external funding.
Availability of data and materials
The datasets used during the current study are available from the correspond-
ing author upon reasonable request.
Declarations
Consent for publication
Not applicable.
Competing of interests
The authors declare no competing interests.
Received: 7 November 2023 Accepted: 8 January 2024
References
1. Abiwinanda N, Hanif M, Hesaputra ST, Handayani A, Mengko TR. Brain
tumor classification using convolutional neural network. IFMBE Proc.
2019;68:183–9.
2. Ayadi W, Elhamzi W, Charfi I, Atri M. Deep CNN for brain tumor classifica-
tion. Neural Process Lett. 2021;53:671–700.
3. Badža MM, Barjaktarović MČ. Classification of brain tumors from MRI
images using a convolutional neural network. Appl Sci. 1999;2020:10.
4. Saravanan S, Kumar VV, Sarveshwaran V, Indirajithu A, Elangovan D, Alla-
year SM. Computational and Mathematical Methods in Medicine Glioma
Brain Tumor Detection and Classification Using Convolutional Neural
Network. Comput Math Methods Med. 2022;2022:4380901.
5. Ge C, Gu IYH, Jakola AS, Yang J. Deep semi-supervised learning for brain
tumor classification. BMC Med Imaging. 2020;20:87.
6. Cinar A, Yildirim M. Detection of tumors on brain MRI images using the
hybrid convolutional neural network architecture. Med Hypotheses.
2020;139: 109684.
7. Pei L, Vidyaratne L, Rahman MM, Iftekharuddin KM. Context aware deep
learning for brain tumor segmentation, subtype classification, and sur-
vival prediction using radiology images. Sci Rep. 2020;10:19726.
8. Balasooriya NM, Nawarathna R.D. A sophisticated convolutional neural
network model for brain tumor classification. In Proceedings of the 2017
IEEE International Conference on Industrial and Information Systems,
Peradeniya, Sri Lanka, 15–16 2017; pp. 1–5
Table 11 Comparison of the proposed model with existing studies
Author Year Datasets Method Classification of
Accuracy (%)
Ahmed Wasif Reza [12] 2023 Kaggle, Figshare VGG-16 architecture 96.70
Mahmoud Khaled Abd-Ellah [13] 2018 RIDER, REMBRANDT, and BraTS ECOC-SVM 97.98
Takowa Rahman [27] 2023 Kaggle, REMBRANDT CNN+6 pretrained models 97.12
Anand Deshpande [17] 2021 RIDER and BraTS Discrete cosine transform-based
image fusion combined with CNN
98.14
Proposed model 2023 Figshare, REMBRANDT, TCGA-LGG, TCIA Hybrid CNN 98.56

of 21
9. Deepak VK, Sarath R. Multi-Class Brain Cancer Classification Using Deep
Learning Convolutional Neural Network. PalArch’s J Archaeol Egypt/Egyp-
tol. 2020;17:5341–60.
10. Khan HA, Jue W, Mushtaq M, Mushtaq MU. Brain tumor classification
in MRI image using convolutional neural network. Math Biosci Eng.
2020;17:6203–16.
11. Islam R, Imran S, Ashikuzzaman M, Khan MMA. Detection and Classifica-
tion of Brain Tumor Based on Multilevel Segmentation with Convolu-
tional Neural Network. Biomed Sci Eng. 2020;13:45–53.
12. Reza AW, Hossain MS, Wardiful MA, Farzana M, Ahmad S, Alam F, Nandi
RN, Siddique N. A CNN-Based Strategy to Classify MRI-Based Brain Tumors
Using Deep Convolutional Network. Appl Sci. 2023;13:312.
13. Abd-Ellah MK, Awad AI, Khalaf AA, Hamed HF. Two-phase multi-model
automatic brain tumor diagnosis system from magnetic resonance
images using convolutional neural networks. EURASIP J Image Video
Process. 2018;97:1–10.
14. Noreen N, Palaniappan S, Qayyum A, Ahmad I, Alassafi MO. Brain Tumor
Classification Based on Fine-Tuned Models and the Ensemble Method.
Comput Mater Contin. 2023;67:3967–82. https://doi.org/10.32604/cmc.
2021.014158.
15. Mohsen H, El-Dahshan ESA, El-Horbaty ESM, Salem ABM. Classification
using deep learning neural networks for brain tumors. Future Comput
Inform J. 2018;3:68–71.
16. Chattopadhyay A, Maitra M. MRI-based brain tumor image detection
using CNN based deep learning method. Neurosci Inform. 2022;2:100060.
17. Deshpande A, Estrela VV, Patavardhan P. The DCT-CNN-ResNet50 architec-
ture to classify brain tumors with super-resolution, convolutional neural
network, and the ResNet50. Neurosci Inform. 2021;1:100013.
18. Wankhede DS, Selvaran R. Dynamic architecture based deep learning
approach for glioblastoma brain tumor survival prediction. Neurosci
Inform. 2022;2:100062.
19. Rai HM, Chatterjee K. Detection of brain abnormality by a novel
Lu-Net deep neural CNN model from MR images. Mach Learn Appl.
2020;2:100004.
20. Ramya M, Kirupa G, Rama A. Brain tumor classification of magnetic
resonance images using a novel CNN-based medical image analysis
and detection network in comparison with AlexNet. J Popul Ther Clin
Pharmacol. 2022;29:e97–108. https://doi.org/10.47750/jptcp.2022.898.
21. Bingol H, Alatas B. Classification of Brain Tumor Images using Deep Learn-
ing Methods. Turk J Sci Technol. 2021;16:137–43.
22. Acquarelli J, van Laarhoven T, Postma GJ, Jansen JJ, Rijpma A, van Asten S,
Heerschap A, Buydens LMC, Marchiori E. Convolutional neural networks
to predict brain tumor grades and Alzheimer’s disease with MR spectro-
scopic imaging data. PLoS ONE. 2021;7:e0268881.
23. Bhandari A, Koppen J, Agzarian M. Convolutional neural networks for
brain tumor segmentation. Insights Imaging. 2020;11:1–9.
24. Shelatkar T, Urvashi D, Shorfuzzaman M, Alsufyani A, Lakshmanna K.
Diagnosis of Brain Tumor Using Light Weight Deep Learning Model with
Fine-Tuning Approach. Comput Math Methods Med. 2022;2022:285884.
25. Shwetha V, Madhavi CR, Nagendra KM. Classification of Brain Tumors
Using Hybridized Convolutional Neural Network in Brain MRI images. Int J
Circuits Syst Signal Process. 2022;16:561–70.
26. Anagun Y. Smart brain tumor diagnosis system utilizing deep convolu-
tional neural networks. Multimed Tools Appl. 2023;675:1–27.
27. Rahman T, Islam MS. MRI brain tumor detection and classification using
parallel deep convolutional neural networks. Meas Sens. 2023;26:100694.
28. Al-Zoghby AM, Al-Awadly EMK, Moawad A, Yehia N, Ebada AI. Dual Deep
CNN for Tumor Brain Classification. Diagnostics. 2023;13:2050.
29. Saeedi S, Rezayi S, Keshavarz HR, Niakan Kalhori S. MRI-based brain tumor
detection using convolutional deep learning methods and chosen
machine learning techniques. BMC Med Inform Decis Mak. 2023;23:1–17.
30. Srinivasan S, Bai PSM, Mathivanan SK, Muthukumaran V, Babu JC, Vilce-
kova L. Grade Classification of Tumors from Brain Magnetic Resonance
Images Using a Deep Learning Technique. Diagnostics. 2023;13:1153.
31. Yin Z, Gao H, Gong J, Wang Y. WD-UNeXt: Weight loss function and drop-
out U-Net with ConvNeXt for automatic segmentation of few shot brain
gliomas. IET Image Process. 2023;17:3271–80.
32. Prasad V, Vairamuthu S, Selva Rani B. K-Net-Deep joint segmentation with
Taylor driving training optimization based deep learning for brain tumor
classification using MRI. The Imaging Sci J. 2023. https://doi.org/10.1080/
13682199.2023.2208963.
33. Barboriak D. Data from RIDER_NEURO_MRI. Cancer Imag Arch. 2015.
https://doi.org/10.7937/K9/TCIA.2015.VOSN3HN1.
34. Lisa S, Flanders Adam E, Mikkelsen JR, Tom Andrews DW. Data From
REMBRANDT. Arch: Cancer Imag; 2015. https://doi.org/10.7937/K9/TCIA.
2015.588OZUZB.
35. Pedano N, Flanders AE, Scarpace L, Mikkelsen T, Eschbacher JM, Hermes
B, Ostrom Q. Radiology data from the cancer genome atlas low grade
glioma [TCGA-LGG] collection. Cancer Imag Arch. 2016. https://doi.org/
10.7937/K9/TCIA.2016.L4LTD3TK.
36. Clark K, Vendt B, Smith K, Freymann J, Kirb J, Koppel P, Moore S, Phillips
S, Maffitt D, Pringle M. The cancer imaging archive (TCIA): Maintain-
ing and operating a public information repository. J Digit Imaging.
2013;26:1045–57.
37. Cheng J, Huang W, Cao S, Yang R, Yang W, Yun Z, Wang Z, Feng Q.
Enhanced performance of brain tumor classification via tumor region
augmentation and partition. PLoS ONE. 2015;10: e0140381.
38. Rahim T, Usman MA, Shin SY. A survey on contemporary computer-aided
tumor, polyp, and ulcer detection methods in wireless capsule endos-
copy imaging. Comput Med Imaging Graph. 2020;85: 101767.
39. Rahim T, Hassan SA, Shin SY. A deep convolutional neural network for
the detection of polyps in colonoscopy images. Biomed Signal Process
Control. 2021;68: 102654.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in pub-
lished maps and institutional affiliations.

Hybrid Deep Convolutional Neural Network

More Related Content

Similar to Hybrid Deep Convolutional Neural Network (20)

Recently uploaded (20)

Hybrid Deep Convolutional Neural Network