Measurement: Sensors 26 (2023) 100694

Contents lists available at ScienceDirect

Measurement: Sensors
journal homepage: www.sciencedirect.com/journal/measurement-sensors

MRI brain tumor detection and classification using parallel deep

convolutional neural networks
Takowa Rahman, Md Saiful Islam *
Department of Electronics and Telecommunication Engineering, Chittagong University of Engineering and Technology, Chattogram, Bangladesh

A R T I C L E I N F O A B S T R A C T

Keywords: Convolutional neural network (CNN) is widely used to classify brain tumors with high accuracy. Since CNN
Brain tumor detection collects features randomly without knowing the local and global features and causes overfitting problems, this
Data augmentation research proposes a novel parallel deep convolutional neural network (PDCNN) topology to extract both global
Parallel convolutional neural network
and local features from the two parallel stages and deal with the over-fitting problem by utilizing dropout
ReLU activation function
Softmax function
regularizer alongside batch normalization. To begin, input images are resized and grayscale transformation is
conducted, which helps to reduce complexity. After that, data augmentation has been used to maximize the
number of datasets. The benefits of parallel pathways are provided by combining two simultaneous deep con­
volutional neural networks with two different window sizes, allowing this model to learn local and global in­
formation. Three forms of MRI datasets are used to determine the effectiveness of the proposed method. The
binary tumor identification dataset-I, Figshare dataset-II, and Multiclass Kaggle dataset-III provide accuracy of
97.33%, 97.60%, and 98.12%, respectively. The proposed structure is not only accurate but also efficient, as the
proposed method extracts both low-level and high-level features, improving results compared to state-of-the-art
techniques.

1. Introduction 360people aged 40 and up in the United States are predicted to be

identified with a primary brain tumor. A total of 700,000 Americans are
A brain tumor is a development that develops in the tissues sur­ thought to be affected by primary brain tumors, with benign tumors
rounding the brain or the skull and has a major impact on human life. A accounting for 71% of all instances and malignant tumors accounting for
benign or malignant growth can be characterized in two ways. Primary 29% [6]. Tumors are discovered by conventional medical methods such
cancers begin within the brain, whereas secondary tumors, known as as MRI, and biopsies are taken for examination. MRI is the first medical
brain metastasis tumors, are most commonly generated from tumors procedure to detect cancer [7,8]. Fig. 1 depicts two MRI scans of two
outside the brain. The most prevalent primary brain tumors include different brains.
glioma, meningioma, and pituitary adenomas. Meningioma is a slow- Manually interpreting these photos has become time-consuming,
growing tumor that develops from the brain, spinal cord membrane chaotic, and usually wrong as the number of patients has increased.
layers, and brain. Pituitary adenomas are tumors that emerge in the To alleviate this constraint, a computer-aided diagnostic method that
pituitary gland [1]. These tumors grow unevenly in the brain, putting limits the cost of brain MRI identification must be established. Several
pressure on the existing tissue. Pressure causes a slew of brain issues that efforts have been made to develop a very efficient and reliable approach
have a negative impact on the body [2]. In contrast to benign tumors, for automatically categorizing brain tumors. Handmade qualities are
malignant tumors develop unevenly, causing harm to the surrounding used in traditional machine learning methods, limiting the solution’s
tissues. Several approaches can be used to treat brain tumors. In the durability and raising the cost. However, supervised learning models
treatment of brain tumors, surgical methods are commonly used [3]. can sometimes outperform unsupervised learning techniques, resulting
MRI is chosen over other imaging techniques including computed to­ in an overfitted model that is unsuitable for another huge database.
mography (CT), positron emission tomography (PMT), and X-rays These issues highlight the importance of developing an entirely auto­
because of its non-interfering characteristic [4,5]. By 2022, 72, mated brain tumor categorization based on machine learning. CNN’s

* Corresponding author.
E-mail address: [email protected] (M.S. Islam).

https://doi.org/10.1016/j.measen.2023.100694
Received 22 October 2022; Received in revised form 18 December 2022; Accepted 2 February 2023
Available online 6 February 2023
2665-9174/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
T. Rahman and M.S. Islam Measurement: Sensors 26 (2023) 100694

The precision of their classifications is 94.58%. P. Afshar has proposed

employing capsule networks as a convolutional neural network struc­
ture [18]. Using the Figshare dataset, the suggested scheme focuses on
the primary region of tumors and their connections to neighboring tis­
sues, with a precision of 90.89%. In Ref. [19], Mahmoud Khaled
Abd-Ellah et al. proposed a PDCNNs structure for detecting and classi­
fying gliomas from brain MRI images. The BraTS-2017 dataset is used to
test the suggested PDCNNs. The study uses 1200 photos in the training
phase of the PDCNN, 150 images in the validation phase of the PDCNN,
Fig. 1. Two separate brains are scanned using MRI. Healthy on the right and a and 450 photos in the test phase of the PDCNN. The framework has
tumor on the left [9]. delivered remarkable results in accuracy, sensitivity, and specificity,
(97.44%, 97.0%, and 98.0%, respectively).
design is based on a deep learning model, a neural network that is A CNN is used by A. E. Minarno et al. to detect 3 types of brain tumors
particularly good at picture recognition and classification [10,11]. The on MRI images [20]. In this research, 3264 datasets with detailed photos
goal of this study is to create a fully self-contained PDCNN model for of glioma tumors (926 photos), meningioma tumors (937 photos), pi­
brain tumor categorization using publicly available Kaggle and Figshare tuitary cancers (901 photos), and others with no tumors (500 photos)
datasets [9,12,13]. This study proposes a brain tumor detection and are examined. The CNN approach paired with Hyperparameter Tuning is
classification architecture that comprises two simultaneous deep con­ presented to get optimal results in classifying brain tumor varieties.
volutional neural networks. The PDCNN model is quicker over the Three different scenarios are used to test the model and in the third
conditional random field (CRF)-based techniques because convolutions model evaluation scenario, the outcome shows an accuracy of 96.00%.
are both precise and efficient procedures in terms of time. Batch Using well-known deep transfer learning models such as DenseNet201,
normalization is also utilized in the proposed PDCNN framework to DenseNet169, DenseNet121, MobileNet v2, VGG19, VGG16, and Xcep­
normalize the outcomes of preceding layers. tion [21], focuses on the construction of a unique brain tumor detection
The benefits of parallel pathways are provided by integrating two and categorization system. Using a benchmark dataset, the performance
simultaneous deep convolutional neural networks having two different of the suggested framework is assessed in terms of accuracy and loss.
window sizes, allowing the model to learn global and local features. With DenseNet201, it is noted that training accuracy is 97.49% and
That’s one of the article’s contributions. The work’s main achievements validation accuracy is 96.43% and performs better than all other models
are displayed in three aspects:1) A PDCNN has been created as a fully in terms of accuracy. Using the transfer learning approach in the
autonomous brain tumor detection and classification architecture by InceptionV3 and DenseNet201 models, S.U. Habiba et al. explored a
combining two parallel convolutional neural networks. 2) Boosting the classification problem on a public dataset in Ref. [22]. In order to get a
identification and categorization performance by including high-level good classification outcome and prevent overfitting, data augmentation
and low-level data, and specific brain characteristics. 3) The dynam­ is used. Six convolutional layers that are densely coupled together to
ical properties provided by the two simultaneous convolutional layers extract features from dense layers make up a deep CNN model. This
are considerably improved using feature fusion technology. model’s ability to distinguish between the three different forms of brain
The remaining of this article will be arranged as follows: Section 2 tumors has a classification accuracy of about 96.3%.
contains a synopsis of relevant research and a full evaluation of these In [23], using machine learning approaches, A. K. Budati et al.
studies. The suggested PDCNN model is thoroughly detailed in section 3. describe an automated brain tumor categorization and detection model
A complete comparison of the suggested method to current methodol­ of brain MRI scans. In order to detect the tumor, the preprocessed pic­
ogies, and a discussion of the experimental findings, are included in tures are segmented with a C–V model, followed by GLCM feature
sections 4 and 5. The article closes with section 6, which is the final extraction to extract key features from the segmented images, and
section of the study. classification of the images using the SVM and KNN based on the
extracted feature selection. More accurately than the KNN classifier, the
2. Related work SVM classifier performs at a level exceeding 98%.A novel automated
system for the detection and classification of brain tumors that are
Depending on the classification of brain tumors, there are a variety of broken down into the following categories: MRI image preprocessing,
research papers in the literature. Here are some of the works that have image segmentation, extraction of features, and image categorization
been discussed. has been proposed in Ref. [24]. Utilizing an adaptive filter, the MRI
Habib [14] has suggested a convolutional neural network to detect image noise is removed. The modified K-means clustering algorithm is
brain cancers using the Kaggle binary brain tumor classification used to segment the images, and the gray-level co-occurrence matrix is
dataset-I, used in this article. This approach can achieve an accuracy of employed for feature extraction. Finally, a deep learning model recur­
88.7% using a modified neural network architecture [15]. explains the rent convolutional neural network is utilized to categorize glioma, me­
creation of a model that focuses on an artificial CNN for MRI analysis ningioma, non-tumors, and pituitary tumors with95.17% accuracy.
utilizing mathematical formulas and matrix operations. MRI has been
used to train this neural network on 155 brain tumors and 98 tumor-free 3. Proposed methodology
brains. With a 96.7% accuracy rate in the validation data, the model
accurately indicates the location of a tumor. Anil et al. have suggested a The biggest issue before starting treatment is detecting and classi­
technique [16] that comprises a classification network that splits the fying tumors from brain MRI scans. Although most research on brain
input MRI pictures into two categories: one with a tumor and the other tumor diagnosis has concentrated on tumor slicing and positioning
without one. The classifier in this work is retrained using the transfer techniques, there are few studies on tumor diagnosis as a time-saving
learning method to identify brain cancers. VGG-19 is the most effective, technique. Most researchers are using a convolutional neural network
according to the findings, with an accuracy rate of 95.78%. Muhammad to classify brain tumors but it collects features randomly without
Sajjad et al. have developed a new CNN model to classify brain tumors knowing the global and local features. Since the current approach is
[17]. Segmentation is performed to pick the tumor spot before using inconvenient, time-consuming, and human error-prone, these in­
MRI pictures. In the following stage, the dataset is expanded. The rec­ conveniences display how obligate it is to execute a fully automated
ommended CNN is then used to carry out the categorization procedure. deep learning-based brain tumor classification approach. For brain
tumor identification and categorization employing MRI pictures, a

2
T. Rahman and M.S. Islam Measurement: Sensors 26 (2023) 100694

parallel architecture of two CNNs is presented. Fig. 2 shows a block 3.2. Preprocessing data
diagram of the PDCNN design. The following is the sequence of events
that take place within the suggested structure: the input layer of the Data preprocessing is a technique for cleaning data and preparing it
PDCNNs receives brain MRI pictures, which get preprocessed to reduce for use in a machine-learning model, improving the model’s accuracy
computing complexity. For training purposes, the input images are and efficiency. All the brain images in the MRI datasets do not have the
transformed to 32 × 32 pixels from a variety of pixel heights and widths. same width, height, or size; to provide uniformity for training reasons,
These input photos are converted to grayscale, which helps to reduce all of the photos are scaled to 32 × 32 pixels. These input photos are
complexity. After that, data augmentation is used to create new photos converted to grayscale, which helps to reduce complexity.
from old ones. To train the proposed network, the data set is separated
into training and validation. The PDCNN structure is then used to clas­ 3.3. Data augmentation
sify the input pictures, which include local, global, merging, and output
routes. At the output path, to carry out the brain tumor categorization At this stage, data augmentation is used to expand the amount of data
process, the softmax function is employed. available by modifying the original image because deep learning re­
quires a large amount of data to learn from. It is possible to improve the
3.1. MRI dataset effectiveness of categorized outcomes by supplementing them. Rotation,
scaling, translation, and filtering are all operations that can be applied to
Three different public dataset of brain MRI pictures are used in this images. The filtering procedure is used as a supplement in this article.
research. For simplicity, the Kaggle platform has been used to obtain the The inappropriate information consists of noise being present in all
first publicly attainable dataset of binary-class brain MRI pictures; in this MRI brain pictures, resulting in a poor recognition rate. To provide
paper, this data is referred to as dataset-I [9]. There are 253 brain MRI valuable information, it is required to reduce the noise and unwanted
pictures in this collection, including 98 tumors and 155 non-tumor areas. High-frequency noise is seen in MRI pictures, which is generally
groups. This study makes use of the Figshare dataset of 233 patients’ minimized using a filtering process. The anisotropic diffusion filter is a
brain MRI pictures [12]. These brain MRI pictures are taken at two technique for autonomous noise suppression while maintaining image
Chinese hospitals (Nanfang Hospital and General Hospital). It contains edges. This filter can be used to remove noise from digital photos while
3064 MRI scans of the brain(1426 glioma tumors, 708 meningioma tu­ avoiding edge blurring. Table 2 displays the performance of MRI brain
mors, and 930 pituitary tumors); this dataset is identified as dataset-II. images after using an anisotropic diffusion filter.
Finally, the Kaggle website is also used to obtain the other dataset
used in this research [13]; it includes 826, 822, 395, and 827 brain MRI 3.4. Developed Parallel deep convolutional neural networks design
pictures, respectively, of glioma tumor, meningioma tumor, no tumor,
and pituitary tumor. In this study, this data is referred to as dataset-III. Extraction of features and categorization are the two components of
Table 1 displays the different types of brain MRI pictures found across all a standard CNN. The CNN models include input, convolutional, pooling,
datasets. fully connected, and classification layers in their construction, shown in
Fig. 3. For classification, fully connected layers and categorization layers
are utilized, whereas the characteristics are obtained using convolu­
tional and pooling layers.
CNN has been widely used in image/video recognition and classifi­
cation in recent years. CNN can extract image features that are either
global or local, or both, automatically from input images.
This research proposes a new network topology for brain tumor
detection and characterization that consists of two deep convolutional
neural networks operating simultaneously which are shown in Fig. 4.
The PDCNN structure is used to classify the input pictures, which include
local, global, merging, and output routes. To acquire local and global
features, the local and global pathways are used, respectively. In the
local path, the convolutional layers utilize the modest window size of 5
× 5 pixels to offer low-level information with in the pictures. The con­
volutional layers of the global pathway, on the other hand, have a
massive number of filters of 12 × 12 pixels. For each path that down­
samples the convolutional layer output, the max-pooling layer is utilized
after every convolutional layer. The two pathways are joined by a fusion
layer, which creates a single path along with a cascaded link until it
entered the ultimate goal. A batch normalization layer precedes a ReLU
layer, which is followed by two fully linked layers that are coupled to a
dropout layer in the merging route. To deal with the over-fitting prob­
lem, a dropout of 0.3 is applied in the initial layers of the model, which
decreases as the network grows deeper. At the output path, to carry out
the brain tumor categorization process, the softmax function is
employed.
In the PDCNN framework, both global and local features obtained
from the two parallel stages are incorporated. Dropout is a regulariza­
tion strategy for preventing overfitting in training data.
The layers used in the proposed PDCNN framework are discussed
below-

A. Convolutional Layers The convolutional layer is the most signifi­

Fig. 2. The framework of the proposed approach. cant component of the recommended PDCNN framework. Aside from

3
T. Rahman and M.S. Islam Measurement: Sensors 26 (2023) 100694

Table 1
Categorization of MRI brain pictures.
No Tumor Glioma Tumor Meningioma Tumor Pituitary Tumor

MRI Brain Pictures [13]

Table 2
Performance of MRI brain images after using an anisotropic diffusion filter.
No Tumor Glioma Tumor Meningioma Tumor Pituitary Tumor

MRI Brain Pictures [13]

Anisotropic Diffusion Filtered Pictures

D. Integrating Path The fusion procedure’s purpose is to integrate

features from all channels and encrypt inter-group knowledge into a
reduced feature map. Batch normalization and subsequently ReLU
activation follow a merging procedure. The resulting feature map of
an integrating method is determined by equation (2).
Z = σ (BN(f (X))) (2)

Where σ refers the ReLU activation function, BN defines the batch

normalization function, andthe merged extracted features from each
layer in the prior routes are represented by f(X).

E. Regularization As shown in equation (3), to reduce variations and

regulate the transferred data to the kernel function, batch normali­
zation is performed. PDCNN training is aided by BN, which facilitates
weight initialization by providing more feasible activation functions,
simplifying the optimization, achieving optimal results faster, and
generating adjustable gradients.

Fig. 3. The normal CNN model.

m − E[m]
m
̂ = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (3)
Var[m]
the initial layer, the convolutional layer uses maps from the previous
where m̂ denotes normalized activations, m denotes the layer input, E
layer. It generates feature maps as a result.
[m] represents the expectation, and Var[m] is the unbiased variance
B. Max Pooling Layers The input is down-sampled via max-pooling by
estimate.
reducing its dimension. The number of learned parameters should be
Another regularization strategy is a dropout, which ignores neurons
reduced as a result of this procedure, lowering the computing cost
that are picked at random and ceases updating their weights during
and helping avoid overfitting.
training [19].
C. ReLU Layers In the hidden layers, ReLU layers are commonly used.
Details on the PDCNN model that has been proposed, are represented
If the inputs are greater or exceed 0, the outcomes (x, y) are similar;
in Table 3.
alternatively, the outcomes are equivalent to 0, as in equation (1).
fy = max(x, 0) (1)

4
T. Rahman and M.S. Islam Measurement: Sensors 26 (2023) 100694

Fig. 4. The PDCNN model contains four stages in its overall structure: local, global, merging, and output stages in order to categorize the input image.

4. Implementation and evaluation model is shown in Fig. 6 for multi-class Figshare dataset- II for many
ratios. The number of epochs, iterations, elapsed time, and accuracy is
The PDCNN’s implementation code is executed in MATLAB. The calculated for each ratio. At a 90:10 training-testing ratio, the maximum
computer has a 3.2 GHz Intel Core i5 processor, 8 GB of RAM, and a accuracy achieved is 96.10%. This result is achieved after 55 epochs and
Windows operating system installed. 1155 iterations, with a total execution time of 4169 s. A change in the
training-testing ratio causes a change in accuracy. The accuracy of the
PDCNN model improves after adding augmentation to the multi-class
4.1. Classification results Figshare dataset- II. The PDCNN model’s accuracy rises to 97.60%
with data augmentation when the training-testing ratio is 90:10. The
The proposed two simultaneous deep CNN architecture for tumor bold numbers represent the greatest possible outcome.
identification and categorization is validated using three datasets: Fig. 7 shows the efficiency of brain tumor classification using the
dataset-I [9], dataset-II [12], and dataset-III [13]. As shown in Figs. 5–7, PDCNN model over several ratios for the multi-class Kaggle dataset- III.
multiple ratios are provided for training and testing of the proposed The number of epochs, iterations, elapsed time, and accuracy is calcu­
design, including (90:10), (80:20), (70:30), (60:40), and (50:50). lated for each ratio. At an 80:20 training-testing ratio, the greatest ac­
The accuracy of brain tumor classification using the PDCNN model curacy achieved is 95.60%. This result is obtained after 50 epochs and
for multiple ratios is shown in Fig. 5 for binary dataset-I. The number of 800 iterations, with a total execution time of 3465 s. A change in the
epochs, iterations, elapsed time, and accuracy is calculated for each training-testing ratio causes a change in accuracy. The accuracy of the
ratio. At a 90:10 training-testing ratio, the maximum accuracy achieved PDCNN model improves after adding augmentation to the multi-class
is 96.00%. This result is obtained after 90 epochs and 90 iterations, with Kaggle dataset- III. The PDCNN model’s accuracy rises to 98.12
a total execution time of 361 s. A change in the training-testing ratio percent from 95.60% utilizing data augmentation when the training-
causes a change in accuracy. The accuracy of the PDCNN model im­ testing ratio is 80:20. The bold numbers represent the greatest
proves after applying augmentation to the binary classification dataset-I. possible outcome.
When the training-testing ratio is 90:10, the PDCNN model’s accuracy
climbs to 97.33% with data augmentation from 96%. The bold numbers
represent the greatest possible outcome.
The performance of brain tumor categorization using the PDCNN

5
T. Rahman and M.S. Islam Measurement: Sensors 26 (2023) 100694

Table 3
Details on the PDCNN model that has been proposed.
Layer No. Layer Type Properties Activations Learnable Total Learnable

1 Image Input 32 × 32 × 1 images with ‘zero center’ normalization 32 × 32 × 1 – 0

2 Convolutional 128 5 × 5 convolutions with stride [2,2] and padding ‘same’ 16 × 16 × 128 Weights: 5 × 5 × 1 × 128 3328
Bias: 1 × 1 × 128
3 ReLU ReLU 16 × 16 × 128 – 0
4 Cross Channel Cross-channel normalization with 5 channels per element 16 × 16 × 128 – 0
Normalization
5 Max 2 × 2 max pooling with stride [22] and padding ‘same’ 8 × 8 × 128 – 0
6 Convolutional 96 5 × 5 convolutions with stride [2,2] and padding ‘same’ 4 × 4 × 96 Weights: 5 × 5 × 128 × 96 307,296
Bias: 1 × 1 × 96
7 Convolutional 128 12 × 12 convolutions with stride [2,2] and padding ‘same’ 16 × 16 × 128 Weights: 12 × 12 × 1 × 128 18,560
Bias: 1 × 1 × 128
8 ReLU ReLU 4 × 4 × 96 – 0
9 Max 2 × 2 max pooling with stride [22] and stride ‘same’ 2 × 2 × 96 – 0
10 Convolutional 96 5 × 5 convolutions with stride [2,2] and padding ‘same’ 1 × 1 × 96 Weights: 5 × 5 × 96 × 96 230,496
Bias: 1 × 1 × 96
11 ReLU ReLU 1 × 1 × 96 – 0
12 Max 2 × 2 max pooling with stride [22] and stride ‘same’ 1 × 1 × 96 – 0
13 ReLU ReLU 16 × 16 × 128 – 0
14 Cross Channel cross channel normalization with 5 channels per element 16 × 16 × 128 – 0
Normalization
15 Max 2 × 2 max pooling with stride [22] and stride ‘same’ 8 × 8 × 128 – 0
16 Convolutional 96 12 × 12 convolutions with stride [2,2] and padding ‘same’ 4 × 4 × 96 Weights: 12 × 12 × 128 × 96 1,769,568
Bias: 1 × 1 × 96
17 ReLU ReLU 4 × 4 × 96 – 0
18 Max 2 × 2 max pooling with stride [22] and stride ‘same’ 2 × 2 × 96 – 0
19 Convolutional 96 12 × 12 convolutions with stride [2,2] and padding ‘same’ 1 × 1 × 96 Weights: 12 × 12 × 96 × 96 1,327,200
Bias: 1 × 1 × 96
20 ReLU ReLU 1×1× 96 – 0
21 Max 2 × 2 max pooling with stride [22] and stride ‘same’ 1×1× 96 – 0
22 Addition Elements-wise addition of 2 inputs 1×1× 96 – 0
23 Batch Normalization Batch Normalization 1×1× 96 offset: 1 × 1 × 96 192
Scale: 1 × 1 × 96
24 ReLU ReLU 1 × 1 × 96 – 0
25 Fully 512 fully connected layer 1 × 1 × 512 Weights: 512 × 96 49,664
Bias: 512 × 1
26 ReLU ReLU 1 × 1 × 512 – 0
27 Dropout 30% dropout 1 × 1 × 512 – 0
28 Fully 2 fully connected layer 1×1×2 Weights: 2 × 512 1026
Bias: 2 × 1
29 Softmax Softmax 1×1×2 – 0
30 Classification Output – 1×1×2 – 0

Fig. 5. Visualization of PDCNN accuracy on multiple training and testing ratios Fig. 6. Visualization of PDCNN accuracy on multiple training and testing ratios
for binary brain tumor classification dataset-I. for multi-class Figshare dataset- II.

4.2. Performance evaluation

True Positive
Precision =
The outcomes of a categorization system are described using a (True Positive + False Positive)
confusion matrix. The following factor is used to assess performance
True Negative
[25–27]. Recall =
(False Positive + True Negative)
True Positive + True Negative
Accuracy =
(Positive + Negative) 2 ∗ True Positive
F1 Score =
(2 ∗ True Positive + False Positive + False Negative)

6
T. Rahman and M.S. Islam Measurement: Sensors 26 (2023) 100694

Fig. 7. Visualization of PDCNN accuracy on multiple training and testing ratios Fig. 9. Performance parameters comparison of accuracy, precision, recall and
for multi-class Kaggle dataset-III. F1-score with individual models using Figshare dataset- II containing 3 types of
brain tumor images.

Fig. 8 illustrates that the PDCNN model outperforms the standard

CNN model in terms of accuracy, precision, recall, and F1 score. If
augmentation is done to binary classification dataset-I to boost the ef­
ficiency of the suggested PDCNN model, the values of the performance
indicators will grow even more.
Fig. 9 illustrates that the PDCNN model outperforms the standard
CNN model in terms of accuracy, precision, recall, and F1 score. If
augmentation is done to multi-class Figshare dataset-II to boost the ef­
ficiency of the suggested PDCNN model, the values of the performance
indicators will grow even more.
Fig. 10 illustrates that the PDCNN model outperforms the standard
CNN model in terms of accuracy, precision, recall, and F1 score. If
augmentation is done to multi-class Kaggle dataset-III to boost the effi­
ciency of the suggested PDCNN model, the values of the performance
indicators will grow even more.
Fig. 10. Performance parameters comparison of accuracy, precision, recall and
4.3. Error rate calculation F1-score with individual models using Kaggle dataset- III containing 4 types of
brain tumor images.
The error rate is calculated as a percentage of each incorrect pre­
diction divided by the total number of data sets. Accuracy and error rate
are complementary. For example, Accuracy = 1- Error Rate and Error
Rate = 1- Accuracy.
Fig. 11 shows the CNN model with the greatest error rate for the
original dataset. The PDCNN model is used to increase the performance
of the CNN model. Compared to the CNN model, the suggested PDCNN
model has a lower error rate. When augmentation is applied to the
PDCNN model’s input dataset, the error rate is lowered even more and
the model’s performance is significantly improved.
Fig. 12 shows the CNN model with the greatest error rate for the
original dataset. The PDCNN model is used to increase the performance
of the CNN model. Compared to the CNN model, the suggested PDCNN
model has a lower error rate. When augmentation is applied to the

Fig. 11. Training-testing performance vs error rate for binary brain tumor
classification dataset-I.

PDCNN model’s input dataset, the error rate is lowered even more and
the model’s performance is significantly improved.
Fig. 13 shows the CNN model with the greatest error rate for the
original dataset. The PDCNN model is used to increase the performance
of the CNN model. Compared to the CNN model, the suggested PDCNN
model has a lower error rate. When augmentation is applied to the
PDCNN model’s input dataset, the error rate is lowered even more and
the model’s performance is significantly improved.

Fig. 8. Performance parameters comparison of accuracy, precision, recall and

F1-score with individual models using binary classification dataset-I.

7
T. Rahman and M.S. Islam Measurement: Sensors 26 (2023) 100694

instances classified by the classification model matched the data desig­

nated as ground truth after accounting for the arbitrary classifier’s ex­
pected accuracy. The PDCNN model’s kappa is higher than the CNN
model’s kappa. After applying augmentation to the image data of both
the CNN and PDCNN models, the error rate has decreased. However,
after applying augmentation to the input pictures of both the CNN and
PDCNN models, the elapsed time has increased.
Fig. 15 shows that the suggested PDCNN model’s accuracy improves
after augmentation when compared to the original photos. When data
augmentation is included, the proposed model’s accuracy rises to
97.60% from 96.10%.
Table 6 displays the accuracy, error, time, and kappa values for the
CNN and PDCNN models using the original dataset-III and enhanced
Fig. 12. Training-testing performance vs error rate for multi-class Figshare dataset-III, respectively. The PDCNN model’s kappa is higher than the
dataset- II. CNN model’s kappa. After applying augmentation to the image data of
both the CNN and PDCNN models, the error rate has decreased. How­
ever, after applying augmentation to the input pictures of both the CNN
and PDCNN models, the elapsed time has increased.
Fig. 16 shows that the suggested PDCNN model’s accuracy improves
after augmentation when compared to the original photos. When data
augmentation is included, the proposed model’s accuracy rises to
98.10% from 95.60%.

4.5. Comparison of the suggested model to previous efforts on the subject

At the conclusion of the suggested method’s validation, a detailed

comparison is made. Table 7 represents a quick comparison.

5. Discussion

Manually evaluating MRI pictures has become time-consuming,

Fig. 13. Training-testing performance vs error rate for multi-class Kaggle chaotic, and usually erroneous as the number of patients has
dataset- III. increased. Handmade attributes are used in traditional machine learning
methods, limiting the solution’s endurance and increasing the cost.
4.4. Result analysis However, supervised learning models can sometimes outperform unsu­
pervised learning techniques, resulting in an overfitted model that is
The suggested PDCNN approach is presented in full in this section. unsuitable for another huge database. The importance of developing an
Table 4 displays the accuracy, error, time, and kappa values for the entirely automated brain tumor categorization based on machine
CNN and PDCNN models using the original dataset-I and enhanced learning is highlighted by these issues. This research offers a new
dataset-I, respectively. The kappa statistic is a measure of how well the method for brain tumor identification and categorization that consists of
instances classified by the classification model matched the data desig­ two simultaneous deep convolutional neural networks. Local, global,
nated as ground truth after accounting for the arbitrary classifier’s ex­ merging, and output pathways are all part of the PDCNN framework.
pected accuracy. The PDCNN model’s kappa is higher than the CNN Grayscale conversion is applied to the input photos, which helps to
model’s kappa. After applying augmentation to the image data of both simplify the process. Then, using data augmentation, new photos are
the CNN and PDCNN models, the error rate has decreased. However, created from existing ones. The convolutional layers in the local path
after applying augmentation to the input pictures of both the CNN and give local information to the images by using a modest window size of 5
PDCNN models, the elapsed time has increased. × 5 pixels. The convolutional layers of the global path, on the other
Fig. 14 shows that the suggested PDCNN model’s accuracy improves hand, use a massive window size of 12 × 12 pixels. For each path that
after augmentation when compared to the original photos. When data downsamples the convolutional layer output, the max-pooling layer is
augmentation is included, the proposed model’s accuracy rises to utilized after every convolutional layer. The two pathways are joined by
97.30% from 96.00%. a fusion layer, which creates a single path along with a cascaded link
Table 5 displays the accuracy, error, time, and kappa values for the until it entered the ultimate goal. A batch normalization layer precedes a
CNN and PDCNN models using the original dataset-II and enhanced ReLU layer, which is followed by two fully linked layers that are coupled
dataset-II, respectively. The kappa statistic is a measure of how well the to a dropout layer in the merging route. At the output path, to carry out
the brain tumor categorization process, the softmax function is
employed. Dropout is a regularization technique that is also used to
Table 4 prevent overfitting of the training data.
Suggested tumor type categorization outcomes using binary classification
Using three datasets, the technique delivers very accurate brain
dataset-I.
tumor categorization. Each dataset’s accuracy, precision, recall, and F1
Method Original Augmented Measures score are determined. The performance of brain tumor classification
Dataset I Dataset I
Accuracy Error Time Kappa using the PDCNN model for multiple ratios is shown in Figs. 5–7 for
(%) (%) (s) three types of datasets. For datasets I and II, the maximum accuracy is
CNN ✓ 88.24 11.76 37 0.749 achieved when the training-testing ratio is 90:10.; however, for dataset-
✓ 97.33 2.67 121 0.944 III, the best accuracy is achieved when the training-testing ratio is 80:20.
PDCNN ✓ 96.00 4.00 361 0.918 A change in the training-testing ratio causes a change in accuracy. The
97.33 2.67 861 0.944
accuracy of the PDCNN model improves after applying augmentation to
✓

8
T. Rahman and M.S. Islam Measurement: Sensors 26 (2023) 100694

Fig. 14. Confusion matrix of PDCNN using binary classification dataset-I (a) confusion matrix of original dataset-based classification (b)confusion matrix of the
original dataset with augmentation-based classification.

seen in Figs. 11–13. The PDCNN model is used to increase the perfor­
Table 5
mance of the CNN model. Compared to the CNN model, the suggested
Suggested tumor type categorization outcomes using multiclass Figshare data­
PDCNN model has a lower error rate. When augmentation is applied to
set-II.
the PDCNN model’s input dataset, the error rate is lowered even more
Method Original Augmented Measures
and the model’s performance is improved.
Dataset II Dataset II
Accuracy Error Time Kappa The results in Figs. 14–16 show that augmentation improves the
(%) (%) (s) accuracy of the proposed PDCNN model when compared to the original
CNN ✓ 95.80 4.20 833 0.933 images. At the conclusion of the suggested method’s validation, a
✓ 97.90 2.10 1627 0.967
PDCNN ✓ 96.10 3.90 4169 0.938
✓ 97.60 2.40 5803 0.962 Table 6
Suggested tumor type categorization outcomes using multiclass Kaggle dataset-
III.
all three types of datasets. The bold numbers represent the greatest Method Original Augmented Measures
possible outcome. Dataset III Dataset III
Accuracy Error Time Kappa
Figs. 8–10 indicate that the PDCNN model outperforms the standard (%) (%) (s)
CNN model in terms of accuracy, precision, recall, and F1 score. When
CNN 94.10 5.90 629 0.919
augmentation is used on three different types of datasets to improve the
✓
✓ 97.70 2.30 783 0.968
performance of the proposed PDCNN model, the performance metrics PDCNN ✓ 95.60 4.40 3465 0.940
improve even more. ✓ 98.12 1.88 6430 0.974
The original dataset’s CNN model has the maximum error rate, as

Fig. 15. Confusion matrix of PDCNN using Figshare dataset-II (a) confusion matrix of the original dataset-based classification (b)confusion matrix of the original
dataset with augmentation-based classification.

9
T. Rahman and M.S. Islam Measurement: Sensors 26 (2023) 100694

Fig. 16. Confusion matrix of PDCNN using Kaggle dataset-III (a) confusion matrix of the original dataset-based classification (b)confusion matrix of the original
dataset with augmentation-based classification.

PDCNN framework for identifying and categorizing brain tumors from

Table 7
MRI pictures. On the Kaggle dataset, the suggested PDCNN is tested for
Comparison with existing techniques those used Kaggle and Figshare datasets.
binary classification and multi-class classification, with four types of
No Authors Methodology Year Dataset Accuracy tumor pictures in the Kaggle dataset and three types of tumor pictures in
1. P. Afshar et al. Capsule Networks 2019 Figshare 90.89% the Figshare dataset. For three different types of brain tumor datasets,
[18] Dataset-II the structure delivered remarkable accuracy, precision, recall, and F1
2. Chirodip Lodh CNN 2020 Binary 96.08%
score. The proposed structure will be expanded in the future to use a 3D
Choudhury Dataset-I
et al. [26] structure for tumor identification using 3D brain MRI pictures.
3. H. H. Sultan Resize + 2019 Figshare 96.13%
et al. [25] Augmentation + Dataset-II
CNN + Declaration of competing interest
Hyperparameter
Tuning
The authors declare that they have no known competing financial
4. Suhib et al. Gray + Resize + 2020 Binary 96.7%
[15] Flatten + CNN Dataset-I interests or personal relationships that could have appeared to influence
5. Agus Eko Resize + 2021 Kaggle 96.00% the work reported in this paper.
Minarno et al. Augmentation + Dataset-III
[20] CNN +
Data availability
Hyperparameter
Tuning
6. Priyansh et al. CNN Based Transfer 2021 Binary Resnet-50- Data will be made available on request.
[29] Learning Approach Dataset-I 95%,
VGG-16-
Acknowledgement
90%,
Inception-
V3-55% The Chittagong University of Engineering and Technology is funding
7. T.Rahmanet al. Resize + Gray + 2022 Binary 96.9% the endeavor under the project CUET/DRE/2020-21/ETE/009.
[30] Augmen -tation + Dataset-I
Binary + CNN
8. Proposed Resize þ Gray - Binary 97.33% References
Structure scale conversion Dataset-I
+ Augmentation Figshare 97.60% [1] T.A. Sadoon, M.H.A. Al-Hayani, Deep learning model for glioma, meningioma and
þ PDCNN Dataset-II pituitary classification, Int. J. Adv. Appl. Sci. 10 (1) (March 2021) 88–98.
Kaggle 98.12% [2] W. Grisold, A. Grisold, Cancer around the brain, Neuro-oncol. pract. 1 (1) (2014)
Dataset- 13–21.
III [3] M. Kheirollahi, S. Dashti, Z. Khalaj, F. Nazemroaia, P. Mahzouni, Brain tumors:
special characters for research and banking, Adv. Biomed. Res. (6 January 2015).
[4] J. Lukoff, J. Olmos, Minimizing medical radiation exposure by incorporating a new
radiation “vital sign” into the electronic medical record: quality of care and patient
detailed comparison is carried out. The results show that the proposed safety, Perm. J. 27 (September) (2017).
simultaneous network topology exceeds previously published detection [5] R. Vankdothu, M.A. Hameed, Brain tumor MRI images identification and
and classification methodologies. classification based on the recurrent convolutional neural network, Measurement:
Sensors 24 (2022) 28. September.
[6] [Online]. Available: Quick Brain Tumor Facts |, Quick Brain Tumor Facts”,
6. Conclusion and future work National Brain Tumor Society, 2022.
[7] S. Irsheidat, R. Duwairi, Brain tumor detection using artificial convolutional neural
networks, in: 2020 11th International Conference on Information and
Tumors are a terrible menace to humans because cancerous cells can Communication Systems, ICICS), 2020, pp. 197–203.
attack nearby tissue and spread to other parts of the body. The signifi­ [8] M. Toğaçar, B. Ergen, Z. Cömert, BrainMRNet: brain tumor detection using
cance of early brain tumor identification in providing appropriate magnetic resonance images with a novel convolutional neural network model,
Med. Hypotheses 134 (2020) 1. January.
medical treatment has long been recognized. This research introduced a
[9] Brain MRI Images for Brain Tumor Detection | Kaggle.

10
T. Rahman and M.S. Islam Measurement: Sensors 26 (2023) 100694

[10] R. Yamashita, M. Nishio, R.K.G. Do, K. Togashi, Convolutional Neural Networks: an [21] P. Rajak, A.S. Jangde, G.P. Gupta, Towards Design of Brain Tumor Detection
Overview and Application in Radiology,” Insights into Imaging, 2018, Framework Using Deep Transfer Learning Techniques,” Convergence of Big Data
pp. 611–629. Technologies and Computational Intelligent Techniques, IGI Global, 2023,
[11] Z. A. Sejuti, M. S. Islam, “An Efficient Method to Classify Brain Tumor using CNN pp. 90–103.
and SVM,” 2021 International Conference on Information and Communication [22] S.U. Habiba, M.K. Islam, L. Nahar, F. Tasnim, M.S. Hossain, K. Andersson, Brain-
Technology for Sustainable Development (ICICT4SD), pp. 259-263, 2021. DeepNet: a deep learning based classifier for brain tumor detection and
[12] Jun Cheng, 2017 Figshare dataset https://figshare.com/articles/brain tumor classification, in: International Conference on Intelligent Computing &
dataset/1512427. Optimization, Springer, 2023, pp. 550–560.
[13] Brain Tumor MRI Dataset | Kaggle. [23] A.K. Budati, R.B. Katta, An automated brain tumor detection and classification
[14] Brain Tumor Detection Using Convolutional Neural Networks | by Mohamed Ali from MRI images using machine learning techniques with IoT, Environ. Dev.
Habib | Medium 25 November, 2021. Sustain. 24 (9) (2022) 10570–10584.
[15] S. Irsheidat, R. Duwairi, Brain tumor detection using artificial convolutional neural [24] R. Vankdothu, M.A. Hameed, Brain tumor MRI images identification and
networks, in: 2020 11th International Conference on Information and classification based on the recurrent convolutional neural network, Measurement:
Communication Systems, ICICS), 2020, pp. 197–203. Sensors 24 (December 2022).
[16] A. Anil, A. Raj, H.A. Sarma, N. Chandran R, P.L. Deepa, Brain tumor detection from [25] H.H. Sultan, N.M. Salim, W. Al-Atabany, Multi-classification of brain tumor images
brain MRI using deep learning, Int. J. Innovat. Res. Appl. Sci. Eng. (IJIRASE) 3 using deep neural network, IEEE Access (2019) 69215–69225. May 27.
(Issue 2) (August 2019) 68–73. [26] D. Lamrani, B. Cherradi, O.E. Gannour, M.A. Bouqentar, L. Bahatti, Brain tumor
[17] M. Sajjad, S. Khan, K. Muhammad, W. Wu, A. Ullah, S.W. Baik, Multi-grade brain detection using MRI images and convolutional neural network, Int. J. Adv.
tumor classification using deep CNN with extensive data augmentation, J. comput. Comput. Sci. Appl. 13 (No. 7) (2022).
sci. 30 (2019) 174–182. [27] A. Nayan, A.N. Mozumder, M.R. Haque, F.H. Sifat, K.R. Mahmud, A.K.A. Azad, M.
[18] P. Afshar, K.N. Plataniotis, A. Mohammadi, Capsule networks for brain tumor G. Kibria, A deep learning approach for brain tumor detection using magnetic
classification based on MRI images and coarse tumor boundaries, in: 2019 IEEE resonance imaging, Int. J. Electr. Comput. Eng. 13 (No. 1) (February 2023)
International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 1039–1047.
2019, pp. 1368–1372. [29] S. Priyansh, A. Maheshwari, S. Maheshwari, Predictive modeling of brain tumor: a
[19] M.K. Abd-Ellah, A.I. Awad, H.F.A. Hamed, A.A.M. Khalaf, "Parallel deep CNN deep learning approach, Innovations Comput. Intell. Comput. Vis. 3 (Issue 2)
structure for glioma detection and classification via brain MRI Images.", in: 2019 (2021) 275–285.
31st International Conference on Microelectronics, ICM), 2019, pp. 304–307. [30] T. Rahman, M.S. Islam, "MRI brain tumor classification using deep convolutional
[20] A.E. Minarno, M.H.C. Mandiri, Y. Munarko, Hariyady,"Convolutional neural neural network,", in: 2022 International Conference On Innovations In Science,
network with hyperparameter tuning for brain tumor classification, Kinetik: Game Engineering And Technology (ICISET), Chittagong, Bangladesh, 2022,
Technol. Inf. Syst. Comput. Network, Comput. Electron. Control 6 (2) (May 2021) pp. 451–456.
127–132.

11