International Journal of Intelligent Networks 5 (2024) 30–37

Contents lists available at ScienceDirect

International Journal of Intelligent Networks

journal homepage: www.keaipublishing.com/en/journals/
international-journal-of-intelligent-networks

Deep CNN based brain tumor detection in intelligent systems

Brij B. Gupta a, c, d, e, **, Akshat Gaurav b, *, Varsha Arya f, g
a
International Center for AI and Cyber Security Research and Innovations, & Department of Computer Science and Information Engineering, Asia University, Taichung
413, Taiwan, China
b
Ronin Institute, Montclair, NJ, USA
c
Department of Electrical and Computer Engineering, Lebanese American University, Beirut, 1102, Lebanon
d
Symbiosis Centre for Information Technology (SCIT), Symbiosis International University, Pune, India
e
School of Computing, Skyline University College, P.O. Box 1797, Sharjah, United Arab Emirates
f
Department of Business Administration, Asia University, Taiwan, China
g
Center for Interdisciplinary Research, University of Petroleum and Energy Studies (UPES), Dehradun, India

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

Keywords: The early detection of brain tumor is crucial for effective treatment and improved patient prognosis in Industrial
Brain tumor detection Information Systems. This research introduces a novel computational model employing a three-layer Convolu-
Deep learning tional Neural Network (CNN) for the identification of brain tumors in Industrial Information Systems. Leveraging
Convolutional neural network (CNN)
advanced computational techniques, this proposed model can autonomously detect intricate patterns and fea-
Medical Imaging, Industrial information
systems
tures from medical imaging data, resulting in more accurate and expedited diagnoses. With an impressive 90 %
precision rate, our model demonstrates the potential to serve as a valuable tool for medical professionals working
in the field of neuroimaging. By presenting a dependable and precise computational model, this study contributes
to the advancement of brain tumor identification within the domain of medical imaging. We anticipate that our
methodology will aid healthcare providers in making more accurate diagnoses, thereby leading to enhanced
patient outcomes. Potential avenues for future research encompass refining the model’s fundamental architec-
ture and exploring real-time therapeutic applications.

1. Introduction emergency situations, particularly highlighted by the recent pandemic

[3,19]. Technological advancements, particularly in the realm of the
Recent progress in healthcare is attributable to a confluence of Internet of Things (IoT) and intelligent devices, have been instrumental
technological innovations, demographic shifts, and the global impact of in addressing the scarcity of healthcare resources [4]. Additionally, a
events such as the COVID-19 pandemic. The imperative for enhanced notable shift in research priorities, transitioning from isolated
data sharing across institutions, coupled with the necessity for large- hospital-centric treatments to integrated community-based preventative
scale data processing and quality enhancement in medical services, approaches, is exemplified in the recent guidelines for diabetic foot ulcer
has catalyzed the exploration of privacy-preserving methodologies management, reflecting a broader trend in healthcare research [5,20].
within healthcare blockchain systems [1]. Concurrently, the expanding Deep learning, a branch of artificial intelligence, has emerged as a
role of music in healthcare, encompassing formalized music therapy, groundbreaking force in healthcare, revolutionizing a wide array of
evidence-based rehabilitation practices, and the engagement of profes- medical practices. Its application in medical imaging is extensive,
sional and community musicians, signifies a broader inclusion and including, but not limited to, pneumonia classification from X-rays and
diversification of therapeutic approaches [2]. detection of calcifications in intravascular ultrasound images [6,7,21].
Moreover, the evolution of smart healthcare systems is largely driven Additionally, deep learning demonstrates significant potential in diag-
by demographic trends such as an aging population and the prevalence nosing neurological disorders through advanced neuroimaging tech-
of chronic diseases, alongside the exigency to efficiently manage niques [8]. Deep learning models also used in orthopedics to detect

* Corresponding author.
** Corresponding author. International Center for AI and Cyber Security Research and Innovations, & Department of Computer Science and Information Engi-
neering, Asia University, Taichung 413, Taiwan, China.
E-mail addresses: [email protected] (B.B. Gupta), [email protected] (A. Gaurav).

https://doi.org/10.1016/j.ijin.2023.12.001
Received 9 November 2023; Received in revised form 14 December 2023; Accepted 22 December 2023
Available online 28 January 2024
2666-6030/© 2024 The Authors. Published by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
B.B. Gupta et al. International Journal of Intelligent Networks 5 (2024) 30–37

Fig. 1. Model architecture.

31
B.B. Gupta et al. International Journal of Intelligent Networks 5 (2024) 30–37

many deformations(“Deep Learning - a Rising Curve in the Field of Or- The total number of trainable parameters in the model is 6,172, making
thopaedics” 2022). it a relatively lightweight neural network. The model’s forward and
In addition to that. deep learning is also used to detect many backward pass size is estimated at 7.74 MB, while the model parameters
advanced-level diseases, such as brain tumors. Due to the decision- occupy a mere 0.02 MB of memory. The model’s total estimated size,
making ability of deep learning, it is used in medical decision-making including input data, is 8.16 MB.
[9–12]. Additionally, the integration of deep learning with the
Internet of Medical Things (IoMT) has improved the decision-making 4. Optimizer
process [13,24,25]. Due to these advantages of deep learning, we used
this in our proposed model. In our study, the selection of the Adam optimizer (Adaptive Moment
Estimation) was a pivotal factor in the optimization of our deep learning
2. Related work models. Adam, renowned for its efficacy in deep learning contexts, in-
tegrates the principles of RMSprop and Momentum optimizers. Due to
According to past research deep learning models are increasingly this, Adam updates proposed model parameters both efficiently and
being used in healthcare prediction. Bhavya et al.[14] highlights the use accurately. The operational framework and efficacy of the Adam opti-
of deep learning models such as LSTM/Bi-LSTM, RNN, CNN, RBM, and mizer are elaborated in Algorithm 1.
GRU on different healthcare applications. Deep learning models’ results
were compared to those of top-tier machine learning models and prog-
nostic scoring systems using publicly accessible healthcare datasets by
Purushotham et al.[15]. When using the ‘raw’ clinical time series data as
input characteristics to the models, they discovered that deep learning
models consistently outperform all the other techniques. To forecast
future healthcare costs on an individual level using claims data, Zeng
et al. [16] suggested a multi-view deep learning architecture. According
to their empirical findings, their suggested strategy is superior to all
baselines in forecasting healthcare costs.
Nguyen et al. present three deep ensemble learning algorithms that
outperform classic ML and DL methods on statistical, image-based, and
sequential benchmark datasets. Ahmedt-Aristizabal et al.[18] discuss
the many network designs and their medical applications, such as
functional connectivity, anatomical structure, and electrical-based
analysis. Abdullah et al. [17] reviews the applications of Bayesian
deep learning in healthcare, including medical imaging tasks, clinical
signal processing, medical natural language processing, and electronic
health records. Mandle et al. [22], proposed CNN based method for
brain tumer detection. Suganya et al,. [23] proposed COVID-19 detec-
tion using chest X-ray. Overall, these papers suggest that deep learning
models, including ensemble learning, graph neural networks,
multi-view deep learning, and Bayesian deep learning, are effective for
healthcare prediction and analysis. 5. Result and discussion

3. Proposed approach 5.1. System information

This section presents the proposed model for brain tumor detection. In our study, we employed the Kaggle dataset and developed our
The model’s architecture (Fig. 1) is summarized as follows: deep learning model within the PyTorch environment. The model
training process was accelerated by utilizing NVIDIA GPUs, specifically
• Input Layer: The model takes as input images with a size of 64x64 Tesla T4 GPUs, with CUDA version 11.4 support. These GPUs facilitated
pixels. The input layer’s shape is 8x2. It indicates that the model parallel processing, significantly enhancing computational performance
processes the data in batches of 8 images at a time. during training. The GPU utilization during training remained low,
• Feature Extraction Layers: The model consists of three sequential indicating efficient resource management. The Tesla T4 GPUs featured
blocks, each comprising two convolutional layers (Conv2d) followed dedicated memory resources and exhibited stable performance, with
by Rectified Linear Unit (ReLU) activation functions. These con- GPU 0 running at 57 ◦ C and consuming 29 W of power, while GPU 1
volutional layers are responsible for extracting features from the operated at 44 ◦ C with minimal power usage (9 W) during our experi-
input images. After each pair of convolutional layers, there is a ments. This GPU configuration enabled us to train our deep learning
MaxPooling2D layer to downsample the spatial dimensions. Dropout model efficiently.
layers are included to prevent overfitting, and Batch Normalization is
applied for improved training stability. 5.2. Data preprocessing
• Flatten Layer: After feature extraction, the Flatten layer reshapes
the output from the convolutional layers into a one-dimensional In our data preprocessing pipeline, we divided our dataset into
tensor. training and testing sets. To facilitate efficient model training, we uti-
• Fully Connected Layers: The last part of the model involves two lized PyTorch’s data loaders, with a total batch size of 32 for batch
fully connected (Linear) layers. These layers learn to map the flat- processing. As a standardization step, we resized all the images to a
tened features to the final output, which has a shape of [8, 2]. The consistent size of 64x64 pixels, ensuring uniformity in image di-
number of parameters in these layers is 1282. mensions. Additionally, we employed data transformations to enhance
the model’s performance. One key transformation involved the use of
In summary, the model consists of a series of convolutional layers for PyTorch’s Random Horizontal Flip transformer with a probability of 0.5.
feature extraction, followed by fully connected layers for classification. This transformation randomly flipped the images horizontally with a 50

32
B.B. Gupta et al. International Journal of Intelligent Networks 5 (2024) 30–37

Fig. 2. Accuracy and loss curve.

Fig. 3. Classification report.

% chance during training, introducing diversity in the dataset and Initially, the training accuracy is 65.28 %, it then gradually increases
helping the model generalize better by learning from both original and with each epoch. The test accuracy started lower, at 41.67 %, but also
horizontally flipped versions of the images. These preprocessing steps improved over time. Simultaneously, the training loss and test loss
collectively contributed to the improved accuracy of our brain tumor decreased with each epoch. Fig. 2 presents the accuracy and loss curves.
detection model.

5.4. Model evaluation

5.3. Accuracy and loss curves
The classification report offers (Fig. 3) an in-depth analysis of our
To train our deep learning model, we ran it for a total of 50 epochs, model’s efficiency.
where every epoch is a full pass through the training dataset. We tracked Here’s an explanation of the key metrics in the classification report:
the model’s accuracy on both the training and test sets during each
iteration. Training loss, training accuracy, test loss, and test accuracy are • Precision: Precision is the degree to which a model is accurate.
all provided for each period. These measures tell us how effectively our • Recall: The model’s recall measures how many instances of a class it
model is able to take what it has learned from the data and apply it to was able to properly identify. When applied to “hemorrhage_data,” a
new situations. recall of 0.80 indicates that 80 % of instances of “hemorrhage_data”

33
B.B. Gupta et al. International Journal of Intelligent Networks 5 (2024) 30–37

Fig. 4. Confusion matrix.

Fig. 5. Comparison of accuracy and loss curves.

34
B.B. Gupta et al. International Journal of Intelligent Networks 5 (2024) 30–37

Fig. 6. Comparison of different models.

are correctly identified by the model. Recall = 1.00 for “non hem- distinguished between “hemorrhage_data” and “non hemorrhage data”
orrhage data,” implying that all “non hemorrhage data” instances are brain tumor pictures by achieving excellent accuracy, recall, and F1-
properly identified by the model. scores for both classes. The model’s durability and efficacy in the job
• F1-score: The F1-score is a balanced measure of a model’s perfor- of brain tumor identification are validated by the high overall accuracy
mance, being the harmonic mean of accuracy and recall. In the of 90 % and consistently good metrics across precision, recall, and F1-
instance of “hemorrhage_data,” the F1-score is 0.89, whereas in the score. These findings lend credence to claims that our approach suc-
case of “non_hemorrhage_data,” it is 0.91. If the F1 score is high, the cessfully classifies photos of brain tumors.
model effectively balances accuracy and recall. The confusion matrix is used for evaluating the performance of the
• Support: Assist illustrates the distribution of samples across cate- proposed model (Fig. 4). It provides a detailed breakdown of how the
gories. Here, we have 10 samples each of “hemorrhage_data” and model’s predictions align with the actual class labels.
“non_hemorrhage_data."
• Accuracy: The model achieves a 90 % overall accuracy, which in- • True Positives (TP): This represents the number of cases where the
dicates that it properly identifies 90 % of the samples. model correctly predicted “hemorrhage_data".
• Macro Average:The global averages for precision, recall, and F1- • False Negatives (FN): This is the number of cases where the model
score are derived by averaging the two groups together. The sce- predicted “non_hemorrhage_data” when the actual class was
nario’s macro average performance metrics (accuracy, recall, and “hemorrhage_data."
F1-score) are all above average at 0.92. • False Positives (FP): This represents the number of cases where the
model incorrectly predicted “hemorrhage_data” when the actual
The classification report shows that our model successfully class was “non hemorrhage data."

35
B.B. Gupta et al. International Journal of Intelligent Networks 5 (2024) 30–37

Table 1 Competing interests

Results comparison.
Model Labels Precision Recall F1-score Accuracy The authors have no relevant financial or non-financial interests to
disclose.
ResNet HD 0.89 0.80 0.84 0.85
NHD 0.82 0.90 0.86 0.85
DenseNet HD 0.78 0.70 0.74 0.75 Data availability
NHD 0.73 0.80 0.76 0.75
ResNet HD 0.89 0.80 0.84 0.85
All data generated or analysed during this study are included in this
NHD 0.82 0.90 0.86 0.85
AlexNet HeD 0.88 0.70 0.78 0.80
published article.
NHD 0.75 0.90 0.82 0.80
Squeeze HD 0.67 0.80 0.73 0.70 Conflict of interests
NHD 0.75 0.60 0.67 0.70
Proposed HD 0.90 0.90 0.90 0.90
NHD 0.9 0.90 0.9 0.90
Authors have no conflicts of interest.

HD= Hemmorhage Data; NHD=Non Hemmorhage Data.

Ethical statement

• True Negatives (TN): This is the number of cases where the model This article does not contain any studies with human participants or
correctly predicted “non hemorrhage data” when the actual class was animals performed by any of the authors.
indeed “non hemorrhage data.” In your matrix, there are 9 true
negative cases. These are successful predictions of brain tumors CRediT authorship contribution statement
without hemorrhage.
Brij B. Gupta: Formal analysis. Akshat Gaurav: Conceptualization.
The evaluation results section provides a detailed breakdown of the Varsha Arya: Data curation.
model’s performance, enabling a nuanced assessment of its strengths
and weaknesses. In this instance, the substantial counts of true positives
Declaration of competing interest
and true negatives indicate the model’s proficiency in accurately cate-
gorizing both brain tumor image types. The limited occurrences of false
☒ The authors declare that they have no known competing financial
positives and false negatives underscore the model’s infrequent mis-
interests or personal relationships that could have appeared to influence
classifications, affirming its efficacy in brain tumor detection.
the work reported in this paper.
5.5. Comparative analysis
References
In this subsection, we evaluated the performance of different models [1] J. Fu, Y. Cai, Privacy-preserving in healthcare blockchain systems based on
(see Fig. 5). The output of these five approaches is presented in Fig. 6 lightweight Message sharing, Sensors 20 (2020) 1898, https://doi.org/10.3390/
and Table 1. The proposed approach achieved a precision of 87 %, recall s20071898.
[2] B. Foster, S. Pearson, A. Berends, C. Mackinnon, The expanding scope, inclusivity,
of 85 %, and F1-score of 0.87 for both classes, resulting in an overall
and integration of music in healthcare: recent developments, research illustration,
accuracy of 87 %. ResNet demonstrated the highest performance, with and future direction, Healthcare 9 (2021) 99, https://doi.org/10.3390/
an accuracy of 85 %, a precision of 0.89 for hemorrhage data and 0.82 healthcare9010099.
[3] X. Zhao, B. Ge, An indicator framework for assessing the readiness of hospital in the
for non-hemorrhage data, and a balanced recall of 0.80 and 0.90,
smart healthcare transformation, IOP Conf. Ser. Earth Environ. Sci. 1101 (2022)
resulting in F1-scores of 0.84 and 0.86, respectively. DenseNet and 072010, https://doi.org/10.1088/1755-1315/1101/7/072010.
AlexNet also showed strong performance with accuracy levels of 75 % [4] D. Sahu, B. Pradhan, A. Khasnobish, S. Verma, D. Kim, K. Pal, The Internet of
and 80 %, respectively. However, SqueezeNet exhibited the lowest ac- Things in geriatric healthcare, J. Healthc. Eng. 2021 (2021) 1–16, https://doi.org/
10.1155/2021/6611366.
curacy at 70 %, with an F1-score of 0.70. These results show that our [5] B. Najafi, R. Mishra, Harnessing digital health technologies to remotely manage
proposed approach performs most efficiently. diabetic foot syndrome: a narrative review, Medicina 57 (2021) 377, https://doi.
org/10.3390/medicina57040377.
[6] C. Singh, Medical imaging using deep learning models, European Journal of
6. Conclusion Engineering and Technology Research 6 (2021) 156–167, https://doi.org/
10.24018/ejeng.2021.6.5.2491.
Our study presents a robust deep learning model for brain tumor [7] H. Sofian, J. Than, S. Mohamad, N. Noor, Calcification detection for intravascular
ultrasound image using direct acyclic graph architecture: pre-trained model for 1-
detection, achieving a commendable accuracy of 90 %. Leveraging a channel image, Indonesian Journal of Electrical Engineering and Computer Science
three-layer Convolutional Neural Network (CNN) and a carefully 22 (2021) 787, https://doi.org/10.11591/ijeecs.v22.i2.pp787-794.
curated dataset, we demonstrated the model’s ability to effectively [8] C. Narmatha, H. Alatawi, H. Alatawi, An analysis of deep learning techniques in
neuroimaging, Journal of Computational Science and Intelligent Technologies 2
differentiate between brain tumor images with and without hemor-
(2021) 7–13, https://doi.org/10.53409/mnaa/jcsit/2102.
rhage. The classification report and confusion matrix highlight the [9] H. Saleh, S. ghany, H. Alyami, W. Alosaimi, Predicting breast cancer based on
model’s strong performance, with high precision, recall, and F1-scores, optimized deep learning approach, Comput. Intell. Neurosci. 2022 (2022) 1–11,
https://doi.org/10.1155/2022/1820777.
as well as a well-balanced distribution of true positives and true nega-
[10] N. Noreen, S. Palaniappan, A. Qayyum, I. Ahmad, M. Imran, M. Shoaib, A deep
tives. These results signify its clinical potential for early diagnosis. Our learning model based on concatenation approach for the diagnosis of brain tumor,
work contributes to the field of medical imaging, offering a reliable tool IEEE Access 8 (2020) 55135–55144, https://doi.org/10.1109/
that could aid healthcare professionals in improving patient outcomes access.2020.2978629.
[11] A. Jaiswal, P. Tiwari, S. Kumar, M. Al-Rakhami, M. Alrashoud, A. Ghoneim, Deep
through timely and accurate brain tumor detection. learning-based smart iot health system for blindness detection using retina images,
IEEE Access 9 (2021) 70606–70615, https://doi.org/10.1109/
Funding statement access.2021.3078241.
[12] Y. Pan, M. Fu, B. Cheng, T. Xin, J. Guo, Enhanced deep learning assisted
convolutional neural network for heart disease prediction on the Internet of
This research work is supported by National Science and Technology medical Things platform, IEEE Access 8 (2020) 189503–189512, https://doi.org/
Council (NSTC), Taiwan Grant No. NSTC112-2221-E− 468-008-MY3. 10.1109/access.2020.3026214.
[13] S. Raoof, M. Durai, A comprehensive review on smart health care: applications,
paradigms, and challenges with case studies, Contrast Media Mol. Imaging (2022)
1–18, https://doi.org/10.1155/2022/4822235.

36
B.B. Gupta et al. International Journal of Intelligent Networks 5 (2024) 30–37

[14] S. Bhavya, Anitha S. Pillai, Prediction models in healthcare using deep learning, in: International Journal on Semantic Web and Information Systems (IJSWIS) 18 (1)
Proceedings Of the 11th International Conference on Soft Computing and Pattern (2022) 1–14.
Recognition (SoCPaR 2019) 11, 195–204, Springer, 2021. [21] B.C. Anil, P. Dayananda, B. Nethravathi, M.S. Raisinghani, Efficient Local Cloud-
[15] Sanjay Purushotham, Chuizheng Meng, Zhengping Che, Yan Liu, Benchmarking Based Solution for Liver Cancer Detection Using Deep Learning, International
deep learning models on large healthcare datasets, J. Biomed. Inf. 83 (2018) Journal of Cloud Applications and Computing (IJCAC) 12 (1) (2022) 1–13.
112–134. [22] A.K. Mandle, S.P. Sahu, G.P. Gupta, CNN-based deep learning technique for the
[16] Xianlong Zeng, Simon Lin, Liu Chang, Multi-view deep learning framework for brain tumor identification and classification in MRI images, International Journal
predicting patient Expenditure in healthcare, IEEE Open Journal of the Computer of Software Science and Computational Intelligence (IJSSCI) 14 (1) (2022) 1–20.
Society 2 (2021) 62–71. [23] G. Suganya, M. Premalatha, S. Geetha, G.J. Chowdary, S. Kadry, Detection of
[17] Abdullah A. Abdullah, Masoud M. Hassan, Yaseen T. Mustafa, A review on COVID-19 cases from chest X-rays using deep learning feature extractor and
bayesian deep learning in healthcare: applications and challenges, IEEE Access 10 multilevel voting classifier, International Journal of Uncertainty, Fuzziness and
(2022) 36538–36562. Knowledge-Based Systems 30 (05) (2022) 773–793.
[18] D. Ahmedt-Aristizabal, M.A. Armin, S. Denman, C. Fookes, L. Petersson, Graph- [24] Arif, Zainab Hussein, Moamin A. Mahmoud, Karrar Hameed Abdulkareem,
based deep learning for medical diagnosis and analysis: past, present and future, Seifedine Kadry, Mazin Abed Mohammed, Mohammed Nasser Al-Mhiqani, Alaa S.
Sensors 21 (2021) 4758. Al-Waisy, and Jan Nedoma. "Adaptive Deep Learning Detection Model for Multi-
[19] S. Li, D. Qin, X. Wu, J. Li, B. Li, W. Han, False alert detection based on deep Foggy Images." International Journal of Interactive Multimedia & Artificial Intelligence,
learning and machine learning, International Journal on Semantic Web and 7 (7), (2022).
Information Systems (IJSWIS) 18 (1) (2022) 1–21. [25] C.L. Stergiou, A.P. Plageras, K.E. Psannis, et al., Secure machine learning scenario
[20] B. Hu, A. Gaurav, C. Choi, A. Almomani, Evaluation and comparative analysis of from big data, cloud computing via internet of things network, in: Handbook of
semantic web-based strategies for enhancing educational system development, Computer Networks and Cyber Security: Principles and Paradigms, 2020,
pp. 525–554.

37