A CNN Approach to Automated Detection and

Classification of Brain Tumors

Md. Zahid Hasan1 , Abdullah Tamim2 , D.M. Asadujjaman3 , Md. Mahfujur Rahman4
Md. Abu Ahnaf Mollick5 , Nosin Anjum Dristi6 , Abdullah - Al - Noman7
1, 2, 7
Dept. of Computer Science & Engineering, University of Rajshahi, Bangladesh
3
Dept. of Computer Science & Engineering, Khulna University of Engineering & Technology, Khulna, Bangladesh
4
Dept. of Computer Science & Engineering, Rajshahi University of Engineering & Technology, Rajshahi, Bangladesh
2, 3, 4, 5, 6
Dept. of Computer Science & Engineering, Varendra University, Rajshahi, Bangladesh
[email protected], [email protected], [email protected], [email protected],
[email protected], [email protected], [email protected]
arXiv:2502.09731v1 [cs.CV] 13 Feb 2025

Abstract—Brain tumors require an assessment to ensure timely approximately 130 unique types of tumors that can arise in
diagnosis and effective patient treatment. Morphological factors the brain and the central nervous system, encompassing both
such as size, location, texture, and variable appearance com- benign and malignant forms. The prevalence of various tumors
plicate tumor inspection. Medical imaging presents challenges,
including noise and incomplete images. This research article differs, with some being exceptionally rare and others often
presents a methodology for processing Magnetic Resonance Imag- observed [4]. Till date approximately 700,000 individuals in
ing (MRI) data, encompassing techniques for image classification the United States are diagnosed with primary brain tumors.
and denoising. The effective use of MRI images allows medical According to research cited in [5], patients between the ages
professionals to detect brain disorders, including tumors. This of 55 and 64 had a 46.1% one-year survival rate, whereas
research aims to categorize healthy brain tissue and brain tumors
by analyzing the provided MRI data. Unlike alternative methods patients between the ages of 65 and 74 had a 29.3% survival
like Computed Tomography (CT), MRI technology offers a more rate. Image segmentation is a technique used to divide a image
detailed representation of internal anatomical components, mak- into several segments, frequently employed in the medical
ing it a suitable option for studying data related to brain tumors. imaging domain. The visual representation of a picture can be
The MRI picture is first subjected to a denoising technique enhanced by extraction for analytical purposes. This transpires
utilizing an Anisotropic diffusion filter. The dataset utilized for
the model’s creation is a publicly accessible and validated Brain when the image is partitioned into multiple distinct segments.
Tumour Classification (MRI) database, comprising 3,264 brain The scientific examination of images predominates in the field
MRI scans. SMOTE was employed for data augmentation and of medical diagnostics. The existence of nuanced differences,
dataset balancing. Convolutional Neural Networks(CNN) such as specific types of noise, and the absence of evidence concerning
ResNet152V2, VGG, ViT, and EfficientNet were employed for impediments in medical imaging complicates the resolution
the classification procedure. EfficientNet attained an accuracy of
98%, the highest recorded. of this issue. However, magnetic resonance imaging (MRI) is
Index Terms—MRI, EfficientNet, Brain Tumor, SMOTE, CNN more advantageous than autonomous computed tomography
(CT) equipment. It emits no radiation and hence has no adverse
effects on the human body. The fundamental elements are the
I. I NTRODUCTION magnetic field and radio waves. It is a widely utilized non-
The brain and the spinal cord are combinedly known as the invasive imaging modality that offers accurate distinction be-
Central Nervous System, which is crucial for the control of nu- tween tissues. The capability of MRI to normalize commonly
merous cellular functions. The functions include organization, affected tissue enhances the imaging of structures of interest
analysis, decision-making, directive issuance, and information in human brain tumors. Researchers have lately encountered
integration [1]. The human brain exhibits extraordinary com- a significant obstacle in the manual segmentation of brain
plexity owing to its distinctive physical architecture. Condi- MRI images [7]. Image segmentation has often been employed
tions such as brain tumors, infections, migraines, and strokes to identify brain tumors. Diverse methodologies necessitate
are a small subset of central nervous system (CNS) disor- a patient-specific training dataset to perform tailored MRI
ders that present considerable challenges in the development, tumor imaging investigations.These types of datasets intensify
diagnosis, and assessment of successful treatment strategies, the difficulties for specialists. And these solutions typically
as noted in [3]. A significant difficulty for radiologists and depend on alternate imaging modalities, such as T1-weighted
neuropathologists is the early detection of brain tumors, which contrast-enhanced images. Various research also shows the use
result from the aberrant proliferation of brain cells. The detec- of Synthetic Minority Oversampling Technique (SMOTE) for
tion of cerebral malignancies by magnetic resonance imaging data augmentation and balancing the dataset [20].
(MRI) is an intricate manual procedure prone to inaccuracies. The rapid growth and unfavorable prognosis of brain tu-
The abnormal proliferation of nerve cells, resulting in the mors, particularly glioblastomas, represent a significant health
creation of a mass, is termed a brain tumor. There are threat. Despite advancements in MRI and CT imaging, nu-
merous cancers are diagnosed late owing to ambiguous symp- ResNEt architectures, attaining accuracies of 84% and
toms. The location of the tumor and the blood-brain barrier 90%.
complicates treatment. Glioblastomas are the most common • In a distinct study [14], researchers proposed two method-
and malignant form of brain cancer, with a median survival ologies for glioma grading, which involved segmentation
of little over one year in advanced stages [26]. Advance- utilizing a customized U-Net model. A regional convo-
ments in neuroimaging, such as functional MRI and Positron lutional neural network (R-CNN) was employed for the
Emission Tomography (PET) scans, enhance tumor detection classification job in each two-dimensional image slice of
[24]. Novel immunotherapy and targeted therapies are under the MRIs. Their proposed 2D Mask R-CNN achieved
investigation [25]. The classification of brain tumors is essen- an accuracy of 96%.Data augmentation enhanced the
tial for resolving these concerns. Histological and molecular outcomes, as evidenced by the classification efficacy of
classifications assist clinicians in predicting tumor behavior the 2D model.
and tailoring treatment options, thereby enhancing outcomes • In the research conducted by A. M. Dikande Simo, two
and minimizing side effects. Brain Tumor classification aids models were proposed, trained utilizing the Brain Tumor
in identifying biomarkers for focused therapy and improved MRI Dataset [27]. Four optimizers were evaluated across
prognostic outcomes. AI methods such as Deep learning three classification tasks, with Adam demonstrating supe-
and Machine Learning enhance classification precision, hence rior performance in differentiating tumor from non-tumor
augmenting diagnosis and treatment [21]–[23]. brains. Where they got 100% training accuracy and 98%
This research categorizes four tumor types: Glioma, No Tu- validation and test accuracy.
mor, Meningioma Tumor, and Pituitary Tumor, utilizing deep • In the research of [28] A. Nag et al. introduces Tumor-
learning techniques to enhance classification precision. The GANet, a sophisticated model that integrates ResNet50
models evaluate MRI scan data to enhance the reliability and and Generative Adversarial Networks (GANs) for the
speed of brain tumor classification, hence aiding in diagnosis classification of brain tumors. The model demonstrates
and treatment planning. exceptional accuracy of 99.53% and achieves precision
and recall rates of 100%, supported by Explainable AI
II. R ELATED W ORKS
methodologies such as LIME. Nevertheless, the depen-
We have assessed multiple previous research efforts related dence on a particular dataset and the restricted exam-
to machine learning-based supervised, semi supervised, and ination of real-world clinical variability may limit its
unsupervised algorithms relevant to time series analysis. We generalizability.
conducted a thorough analysis to identify the shortcomings of • In the research conducted by A. Rath, B. S. P. Mishra, and
the current system. This research improved the classification D. K. Bagal, [29] they uesd pretrained ResNet50 model
framework we developed for brain cancers. to improve accuracy and efficency utilizing a balanced
• Pendela Kanchanamala et al. [9] utilized MRI to de- dataset of 2,577 MRI images having binary class of
veloped an optimization-enhanced hybrid deep learning tumors and healthy instances of patients.
model for classification and detection of brain tumors. A particular application, certain machine learning models
• Emrah Irmak [10] attained an accuracy of 92.66% utiliz- demonstrate superior efficacy compared to others. However,
ing a bespoke CNN model for classifying normal, menin- the effectiveness of these models in classifying cardiovascular
gioma, pituitary, glioma and metastatic brain tumors. illnesses has not yet reached parity. Further advancement is
• For classification of brain cancer Ayadi et al. [11] pro- required to enhance the existing work. Our motivation is to
posed a CNN-based computer assisted diagnosis (CAD) contribute to the field using deep learning methods to improve
method. Three separate datasets were used to conducted the performance and efficiency of detection and classification.
the experiment using 18-weighted layered CNN model.
Where they achieved 94.74% classification accuracy for III. M ETHODOLOGY
brain tumor type classification and for tumor grading, This study involves a systematic approach to analyzing the
they achieved 90.35%. Brain tumor classification (MRI) dataset to perform the most
• Khan et al. [12] (2020) introduced a deep learning accurate classification of brain tumors using various deep-
approach for the classification of brain cancers as ma- learning models. The process is outlined in the accompanying
lignant or benign, utilizing 253 genuine brain MRI scans structural outline, Fig. 1 illustrates the key steps and stages of
supplemented with data augmentation techniques. Edge the analysis.
detection was employed to define the region of interest Fig. 1 illustrates the process of deep learning by using four
in the MRI image before feature extraction using a basic neural networks, namely ResNet152V2, VGG, ViT, and Effi-
CNN model. The achieved categorization accuracy was cientNet model. Every model separately process the supplied
89%. data and produce predictions. The models are subsequently
• The potential of deep learning techniques for glioma compared to determine the most accurate prediction.
classification by MR imaging is examined by Banerjee Utilizing a deep learning methodology, models are devel-
et al. [13]. For 2D images the researchers assessed the oped to calculate accuracy and assess predictions of different
effectiveness of transfer learning employing VGGNet and classes of brain tumors. The models can incorporate the
 Fig. 1. The proposed methodology of the system.

distinctive characteristics and performance measurements that

are pertinent to each position by classifying the tumors. In
addition, the evaluation of the models involves the analysis
of the f1 score, confusion matrix, and Receiver Operating
Characteristic (ROC) curves.
A. Dataset and Experiment
The pipeline begins with feeding the image data into the
system. These input images can be part of any image classi-
fication dataset. This study’s dataset comprises four classes
of various types of tumors, which are used to train and
evaluate the models. The classes are Glioma Tumor, No
Tumor, Meningioma Tumor, and Pituitary Tumor.

TABLE I
D ETAILS OF DATASET

Dataset Class MRI/ Class Dataset

Train Test
No Tumor 938 833 105
Glioma Tumor 941 841 100
Dataset [19]
Meningioma Tumor 929 814 115
Pituitary Tumor 923 849 74
No Tumor 946 841 105
Glioma Tumor 941 841 100
Balanced Dataset
Meningioma Tumor 956 841 115 Fig. 2. The distribution of the original training dataset.
Pituitary Tumor 915 841 74

1) Exploratory Data Analysis: Exploratory Data Analysis

(EDA) is applied to summarize the main features of the data, perspectives and planes. Multiple scans reveal the presence
discover patterns, and identify any potential issues such as of glioma and pituitary tumors, as well as the meningioma
faulty images or outliers. Fig. 2 shows the distribution of the tumor. The aforementioned images have significance in the
training dataset across four different classes of brain tumors. fields of medical diagnosis, treatment planning, and research
Glioma Tumor forms 25.20% of the dataset used for training. about brain tumors.
Furthermore, meningioma tumors constitute 24.39% of the
training sample. The Pituitary Tumor class covers an additional
25.44% of the training data. The No Tumor category comprises
the smallest fraction, accounting for 24.96% of the training
sample.
Fig. 3 illustrates a grid consisting of four MRI brain scans,
each annotated with a specific category of brain tumor: glioma,
pituitary, or meningioma. The images are monochromatic and
depict cross-sections of the human brain acquired from various Fig. 3. Train image data from data augmentation.
 Table I shows that the original training dataset maintains an VGG16, and EfficientNet. The confusion matrix and ROC
equitable distribution of around 850 photos for each type of curves are created to get optimal results and provide a visual
tumor (meningioma, glioma, and pituitary), but has a smaller representation. Next, the models are evaluated using several
sample size (around 800) for the category of ”Meningioma metrics.
Tumor”. The test set shows a comparable trend, with ap- Fig. 4 illustrates the ROC curves of four deep learning
proximately 100 photos allocated to each tumor group and models (VGG16, EfficientNet, ResNet152V2, and ViT) uti-
a somewhat smaller number (roughly 75) for the ”Pituitary lized for brain tumor classification. Models exhibiting AUCs
Tumor” category. approaching 1.00 demonstrate elevated accuracy. ViT and
2) Dataset Preprocessing: A ratio of 80:20 is used to EfficientNet attain impeccable performance (AUC 1.00), al-
divide the training and test datasets. The training dataset is though VGG16 and ResNet152V2, exhibiting AUCs near 1.00,
used to train the model, whereas the test dataset is used demonstrate robust classification with minor discrepancies,
to evaluate the model’s performance during training. Neural indicating greater challenges in differentiating tumor kinds.
networks require inputs of a fixed size, requiring dimensional Fig. 5 presents the confusion matrix, illustrating that Effi-
image adjustment. The size is reliant upon the architecture cientNet shows exceptional classification performance across
of the model being applied. It requires images with exactly all tumor types, particularly in distinguishing glioma from
224x224 pixel size. Proper scaling ensures uniformity across non-tumor cases. The ViT and VGG16 also demonstrate strong
the dataset. Also a well-balanced dataset is necessary for performance, excelling particularly with meningioma tumors.
the successful training of a machine learning model, taking ResNet152V2, however, struggles to differentiate glioma from
into account the presence of a somewhat under-represented the absence of a tumor, leading to the highest rate of misclassi-
”no tumor” class, especially in the training set. Therefore, fications. While all models perform well, EfficientNet stands
the Synthetic Minority Oversampling Technique (SMOTE) is out with the highest precision but ViT and VGG16 exhibit
employed for data augmentation and to balance the dataset difficulties in glioma classification. Conversely, ResNet152V2
by generating synthetic samples. This approach effectively requires further tuning due to its challenges in correctly
addresses class imbalance, enhancing the performance of classifying glioma and no-tumor cases.
machine learning models. Label encoding is a technique used Table III evaluates the efficacy of four models (VGG16,
to convert string labels into numerical values. This includes EfficientNet, ResNet152V2, and ViT) using F1 scores and
one-hot encoding, which converts them into a specific binary accuracy in the classification of brain tumors. EfficientNet
vector representation. In addition, shuffling is implemented surpasses all models, achieving the greatest F1 scores and an
to ensure the random assortment of data, thereby improving overall accuracy of 0.98, exemplifying the optimal equilibrium
the learning skills of the model and reducing the detection of between precision and recall. VGG16 exhibits commendable
organized patterns in the dataset. Before partitioning the data performance, particularly for meningiomas and pituitary tu-
into training and test sets, it is imperative to restructure it. mors, achieving an accuracy of 0.93. ResNet152V2 and ViT
3) Hyperparameters: Hyperparameters in CNNs, including have comparable overall accuracies (0.91); nevertheless, their
layer number, filter size, learning rate, batch size, and epochs performance is somewhat inconsistent, especially in the glioma
influence a model’s performance. The learning rate regulates and no tumor categories. The ViT demonstrates superior
adjustments. The hyperparameters of the experiment are pre- performance in the classification of pituitary tumors.
sented in the table. In summary, EfficientNet is the most dependable and precise
model, with VGG16 closely trailing behind. ResNet152V2 and
TABLE II ViT require enhancement for improved consistency.
M ODELS PARAMETERS FOR INPUT AND CLASSIFICATION STAGE
Table IV illustrates multiple brain tumor classification stud-
Parameters CNN Models ViT ies using machine learning and deep learning models, empha-
Input Shape 224 × 224 72 × 72 sizing this work’s excellent results. The findings show that
No. of epochs 100 100
Batch Size 32 16 modern deep learning architectures outperform conventional
Activation Function Softmax Softmax methods and models. This study shows model performance
Learning Rate 0.0001 0.001 improvements, with EfficientNet obtaining 0.98 accuracy. It
Patch Size - 3
No. of patches - (72/3) × 2 shows its improved brain tumor classification, surpassing
Transformation Layers - 8 previous findings. With an accuracy of 0.93, VGG16 is reliable
for various categorization jobs. ResNet152V2 and ViT score
4) Experimental Setup: This experiment uses 8GB Nvidia 0.91, showing good and consistent performance but slightly
GeForce RTX 4060 and 32 GB RAM. We used NumPy, lower than EfficientNet. However, prior investigations yielded
Sklearn, Matplotlib, Seaborn, Python 3.10, and TensorFlow different results. Study [15] uses classical machine learning
to create pre-trained networks. algorithms like Decision Trees and Naive Bayes. The Deci-
sion Tree performs well with structured data, but the Naive
IV. R ESULT AND D ISCUSSION Bayesian classifier struggles with complex brain tumor data
Deep-learning models are trained using preprocessed data at 0.882. Support Vector Machines (SVMs) with different
and selected classes. This work employs ViT, ResNet152V2, kernel functions achieve 0.97 accuracy in the study [16]. The
 Fig. 4. ROC curves for all models. Fig. 5. Confusion matrix for all models.

TABLE III
P RECISION , R ECALL , AND F1 SCORES FOR ALL MODELS

Classes VGG 16 EfficientNet ResNet152V2 ViT

Precision Recall F1 Precision Recall F1 Precision Recall F1 Precision Recall F1
Glioma tumor 0.92 0.89 0.91 1.00 0.97 0.98 0.97 0.81 0.88 0.90 0.92 0.91
No tumor 0.94 0.89 0.91 0.96 1.00 0.98 0.85 0.90 0.87 0.92 0.81 0.86
Meningioma tumor 0.93 1.00 0.96 0.97 1.00 0.98 0.85 1.00 0.92 0.84 0.92 0.88
Pituitary tumor 0.92 0.98 0.95 1.00 0.98 0.99 0.94 0.97 0.95 0.94 0.98 0.96

work [17] employs MobileNetV2, a deep learning model that of brain tumor classification, demonstrating the capability
has been tuned using a unique metaheuristic known as the of sophisticated deep-learning networks to manage intricate
Contracted Fox Optimization Algorithm (MN-V2/CFO), to medical imaging tasks with exceptional precision.
optimize its hyperparameters. This approach achieves a tumor V. C ONCLUSION
detection accuracy of 97.32%, outperforming the ResNet50
model described in [18], which achieves an accuracy of The primary objective of the project is to categorize MRI
95.14% on pre-processed MRI images. EfficientNet sets a new images using advanced deep-learning models. The efficacy of
standard for brain tumor classification, showing that deep- various deep learning architectures, including ResNet152V2,
learning networks can handle complex medical imaging tasks VGG, ViT, and EfficientNet, is demonstrated. Recent improve-
with precision. ments in MRI image denoising have shown the effectiveness
of hybrid CNN models integrated with anisotropic diffusion
EfficientNet demonstrates superior performance compared filters. These advanced models are meticulously designed
to traditional convolutional neural networks in the classifi- to extract critical features from MRI data. The comparative
cation of brain tumors by utilizing compound scaling. This analysis has clarified the distinct advantages and limitations
approach effectively balances depth, width, and resolution, of each engineering discipline, facilitating informed decision-
thereby improving accuracy while preserving computational making in specific clinical situations. Furthermore, the adept
efficiency. The design is pretrained on extensive datasets, execution of exchange learning has accelerated training and
thereby facilitating improved feature extraction and general- enhanced performance, while EfficientNet has demonstrated
ization. its ability to achieve high accuracy with a remarkable com-
To conclude, the current study exhibits a distinct advance- puting economy. The presented models demonstrate robustness
ment compared to conventional and prior deep learning meth- against fluctuations in picture quality, patient demographics,
ods. EfficientNet establishes a new standard for the accuracy and tumor kinds, indicating their suitability for diverse clinical
 TABLE IV
M ODEL COMPARISON ACROSS STUDIES

Study no. Models used Accuracy F1 Score Precision Recall

Decision Tree 0.96 0.9637 1.00 0.93
[15]
Naive Bayesian 0.882 0.91 0.91 0.91
[16] Support Vector Machine 0.971 - - 0.919
[17] MN-V2/CFO 0.9732 0.8622 0.9768 0.8012
[18] ResNet50 0.9514 0.9515 0.9517 0.9514
VGG16 0.93 0.9325 0.9275 0.94
EfficientNet 0.98 0.9825 0.9825 0.9875
This study
ResNet152V2 0.91 0.9050 0.9025 0.92
ViT 0.91 0.9025 0.90 0.9075

settings. The models achieved an impressive total accuracy [16] J. Amin, M. Sharif, M. Yasmin, and S. L. Fernandes, “A distinctive
of 98%, thereby validating their efficacy. The Contingency approach in brain tumor detection and classification using MRI,” Pattern
Recognit. Lett., vol. 139, pp. 118–127, 2020.
table, marked by a few false positives and false negatives, [17] L. Xu and M. Mohammadi, “Brain tumor diagnosis from MRI based
highlights the model’s capability to accurately differentiate on Mobilenetv2 optimized by contracted fox optimization algorithm,”
between tumor and non-tumor regions. This research has sig- Heliyon, vol. 10, no. 1, p. e23866, 2024.
[18] Q. T. Lu, T. M. Nguyen, and H. L. Lam, “Improving brain tumor MRI
nificantly advanced therapeutic image processing and provides image classification prediction based on fine-tuned MobileNet,” Int. J.
a feasible method to better patient outcomes and refine clinical Adv. Comput. Sci. Appl., vol. 15, no. 1, 2024.
procedures in brain tumor detection. [19] S. Bhuvaji, A. Kadam, P. Bhumkar, S. Dedge, and S. Kanchan, “Brain
tumor classification (MRI).” Kaggle, 2020.
[20] D. Kawahara, A. Jensen, J. Yuan, Y. Nagata, and Y. Watanabe, “Pre-
R EFERENCES dicting the BRAF mutation with pretreatment MRI radiomics features
for melanoma brain metastases receiving Gamma Knife radiosurgery,”
[1] D. Y. Lee, “Roles of mTOR signaling in brain development,” Exp.
Clin. Radiol., vol. 78, no. 12, pp. e934–e940, 2023.
Neurobiol., vol. 24, no. 3, pp. 177–185, 2015.
[21] D. N. Louis et al., “The 2016 world health organization classification of
[2] M. M. Zahoor et al., “A new deep hybrid boosted and ensemble learning-
tumors of the central nervous system: A summary,” Acta Neuropathol.,
based brain tumor analysis using MRI,” Sensors (Basel), vol. 22, no. 7,
vol. 131, no. 6, pp. 803–820, 2016.
p. 2726, 2022.
[22] A. C. Tan, D. M. Ashley, G. Y. López, M. Malinzak, H. S. Friedman, and
[3] M. Arabahmadi, R. Farahbakhsh, and J. Rezazadeh, “Deep learning M. Khasraw, “Management of glioblastoma: State of the art and future
for smart healthcare—A survey on brain tumor detection from medical directions,” CA Cancer J. Clin., vol. 70, no. 4, pp. 299–312, 2020.
imaging,” Sensors (Basel), vol. 22, no. 5, p. 1960, 2022. [23] B. H. Menze et al., “The Multimodal Brain Tumor Image Segmentation
[4] D. V. Gore and V. Deshpande, “Comparative study of various techniques Benchmark (BRATS),” IEEE Trans. Med. Imaging, vol. 34, no. 10, pp.
using deep Learning for brain tumor detection,” in Proceedings of the 1993–2024, 2015.
2020 IEEE International Conference for Emerging Technology (INCET), [24] “Comprehensive cancer information,” Cancer.gov, 01-Jan-1980. [On-
Belgaum, India, 2020, pp. 1–4. line]. Available: https://www.cancer.gov/.
[5] B. B. Yavuz, G. Kanyilmaz, and M. Aktan, “Factors affecting survival in [25] “Clinicaltrials.gov,” Clinicaltrials.gov. [Online]. Available:
glioblastoma patients below and above 65 years of age: A retrospective https://clinicaltrials.gov/.
observational study,” Indian J. Cancer, 2021. [26] “Information and resources about cancer: Breast, colon, lung, prostate,
[6] A. Maksoud, “REAl-Awadi Brain Tumor Segmentation Based on a skin,” Cancer.org. [Online]. Available: https://www.cancer.org/.
Hybrid Clustering Technique,” pp. 71–81, 2015. [27] A. M. Dikande Simo, A. Tchagna Kouanou, V. Monthe, M. Kameni
[7] I. Despotovit, “Wilfried Philips MRI Segmentation of the Human Brain: Nana, and B. Moffo Lonla, “Introducing a deep learning method for
Challenges, Methods, and Applications,” pp. 1–23, 2014. brain tumor classification using MRI data towards better performance,”
[8] A.-S. Capelle, O. Colot, and C. Fernandez-Maloigne, “Evidential seg- Inform Med Unlocked, vol. 44, 2024, doi: 10.1016/j.imu.2023.101423.
mentation scheme of multi-echo MR images for the detection of brain [28] A. Nag et al., “TumorGANet: A transfer learning and generative
tumors using neighborhood information,” Inf. Fusion, vol. 5, no. 3, pp. adversarial network- based data augmentation model for brain tumor
203–216, 2004. classification,” IEEE Access, vol. 12, pp. 103060–103081, 2024.
[9] P. Kanchanamala, Revathi, and M. B. J. Ananth, “Optimization-enabled [29] A. Rath, B. S. P. Mishra, and D. K. Bagal, “ResNet50-based Deep
hybrid deep learning for brain tumor detection and classification from Learning model for accurate brain tumor detection in MRI scans,” Next
MRI,” Biomed. Signal Process. Control, vol. 84, no. 104955, p. 104955, Research, vol. 2, no. 1, p. 100104, 2025.
2023.
[10] E. Irmak, “Multi-classification of brain tumor MRI images using deep
convolutional neural network with fully optimized framework,” Iran. J.
Sci. Technol. Trans. Electr. Eng., vol. 45, no. 3, pp. 1015–1036, 2021.
[11] W. Ayadi, W. Elhamzi, I. Charfi, and M. Atri, “Deep CNN for brain
tumor classification,” Neural Process. Lett., vol. 53, no. 1, pp. 671–700,
2021.
[12] H. Ali Khan et al., “Brain tumor classification in MRI image using
convolutional neural network,” Math. Biosci. Eng., vol. 17, no. 5, pp.
6203–6216, 2020.
[13] S. Banerjee, S. Mitra, F. Masulli, and S. Rovetta, “Deep radiomics
for brain tumor detection and classification from multi-sequence MRI,”
arXiv [cs.CV], 2019.
[14] Y. Zhuge et al., “Automated glioma grading on conventional MRI images
using deep convolutional neural networks,” Med. Phys., vol. 47, no. 7,
pp. 3044–3053, 2020.
[15] J. Naik and S. Patel, “Tumor detection and classification using decision
tree in brain MRI,” International journal of computer science, 2014.