Brain Tumor Classification in Magnetic Resonance Imaging using Convolutional Neural Networks and Transfer Learning⋆

Abstract

Brain tumors pose a significant threat with the potential to disrupt critical brain functions and manifest neurological symptoms, warranting the highest concern. The evaluation of these tumors relies on various imaging methods, including Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and ultrasound. Particularly, MRI of the brain is renowned for its capability to provide vital insights into brain structure and tissue irregularities. This study harnesses the transformative influence of technology, notably artificial intelligence (AI) and deep learning (DL), to address this challenge.The novel approach involves the integration of Convolutional Neural Networks (CNNs) with transfer learning from VGG19 and ResNet. The primary objective is the classification of brain tumors into four distinct categories: meningioma, glioma, pituitary adenoma, and cases without tumors. The CNN model in isolation achieves an impressive 97.23% accuracy rate. However, when integrated with VGG19 and ResNet, the accuracy soars to an even higher 98.26%. This innovative amalgamation of technologies holds immense promise for enhancing the precision of brain tumor classification, potentially reshaping the landscape of neuroimaging and healthcare.

Keywords

• eol>Brain Tumor
• Transfer Learning
• Classification
• CNNN
• VGG19
• ResNEt50

1. Introduction

discussions related to brain tumors, treatment planning, and prognosis, the World Health Organization 1 has deThe human brain serves as the central control hub for vised a classification and grading system. This system a multitude of bodily functions, including motor coordi- categorizes tumors based on the type of cells they consist nation, sensory processing, and vital physiological pro- of or their primary site of origin [ 2 ]. cesses [ 1 ]. Any disruption within the brain, such as the Brain MRI images hold a pivotal role in the detection of emergence of a tumor, has the potential to interfere with tumors and the modeling of their progression, providing its normal operations. essential guidance for treatment decisions. When com

A brain tumor comprises an abnormal cluster of cells pared to alternative imaging techniques such as CT scans within the brain or the cranial cavity. These tumors can or ultrasound, MRI scans ofer a wealth of comprehensive vary widely in nature, ranging from benign to potentially data, enabling the detailed examination of brain structure life-threatening. They are categorized into primary tu- and the precise identification of anomalies within brain mors (originating within the brain) or metastatic tumors tissue [ 3 ]. (originating elsewhere in the body and spreading to the The impact of technology, especially artificial intelbrain). Treatment approaches for these tumors depend ligence (AI) and deep learning (DL), on the field of on factors such as type, size, and location. To facilitate medicine is undeniable, and MRI image processing exemplifies this transformation. Deep Learning (DL), with a 6th International Hybrid Conference On Informatics And Applied Math- special focus on Convolutional Neural Networks (CNNs), ematics, December 6-7, 2023 Guelma, Algeria ofers distinct advantages. These include automated fea* Corresponding author. ture extraction, heightened accuracy in identifying sub†$Thfaetsmeaa.ucthhaohrbsacro@nutrni bivu-tteidareeqt.udazll(yC.. Fatma); tle patterns and irregularities, scalability to handle vast [e-mail] (M. Medjeded); datasets, and the ability to continuously enhance [e-mail] (M. Said); formance through retraining with new data. These [e-mail] (B. Mohamed); tributes position CNNs as powerful tools for the [e-mail] (L. A. Zakaria) © 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License 1https://www.who.int/

Attribution 4.0 International (CC BY 4.0). ing and diagnosis of brain tumors in MRI images [ 4 ]. normal - using CNN. The model attained an accuracy of CNNs represent a class of deep learning models specifi- 98.27%. cally tailored for data structured in grids, such as images. Hossain et al. [ 11 ] proposed a method to extract brain They draw inspiration from the visual processing system tumors from 2D MRI images using Fuzzy C-Means clusin the animal brain, allowing them to preserve spatial in- tering, traditional classifiers, and a convolutional neural formation while capturing local image features [ 5 ]. This network. The experimental study utilized a benchmark method is highly efective, primarily due to its strong fea- dataset (BraTS) with various tumor characteristics. To ture extraction capabilities [ 6 ]. The evolution of CNNs diferentiate between normal and abnormal pixels based has given rise to various architectural models like Resid- on texture and statistical features, CNN achieved a disual Network (ResNet), Network in Network (NiN), VGG, tinction with an accuracy of 97.87%. and GoogleNet [ 7 ].Transfer learning, a technique that Sultan et al. conducted a study in [ 12 ] where they introtransfers knowledge from one domain to another, has duced a DL model based on a CNN for classifying various demonstrated its value across diverse domains, applica- brain tumor types using two publicly available datasets. tions, and data distributions in both research and training The initial dataset classified tumors into meningioma, [ 7 ]. In this paper, we present a comprehensive approach glioma, and pituitary tumors, encompassing 233 cases for classifying brain tumors into four distinct categories: and comprising a total of 3064 T1-weighted Contrastmeningioma, glioma, pituitary adenoma, and cases with- Enhanced (CE) images. The second dataset diferentiated out tumors. Our approach involves the development of among three glioma grades (Grade II, Grade III, and Grade a CNN model, followed by training two transfer learn- IV), involving 73 patients and including 516 T1-weighted ing models—VGG19 and ResNet—on the same dataset. Contrast-Enhanced (CE) images. The proposed network The innovation lies in seamlessly integrating these three structure consists of 16 layers and achieves an accuracy models into a unified framework, with the primary ob- of 96.13% and 98.7% for the two studies, respectively. jective of enhancing the accuracy and precision of brain Similarly, Nayak et al. [ 13 ] introduced a CNN-based tumor categorization in MRI scans. This novel technique dense EficientNet model with min-max normalization holds promise as an alternative for more efective MRI for classifying 3260 T1-weighted contrast-enhanced brain diagnosis and treatment planning. MRI images collected from Kaggle into four categories( glioma, meningioma, pituitary, and no tumor). The experimental results demonstrated performance, with a 2. Related Work training accuracy of 99.97% and a testing accuracy of 98.78%.

The process of manually identifying and categorizing Khan et al. [ 14 ] introduced an automated brain tumor brain tumors in large databases of medical images during classification system that employs two DL models. The routine clinical tasks incurs substantial costs in terms of system is designed to classify brain tumors into binary efort and time. Contemporary solutions that leverage categories (normal and abnormal) using a publicly availMachine Learning (ML) and Deep Learning (DL) method- able CE-MRI dataset consisting of 3064 MRI images. Adologies for brain tumor segmentation, detection, and clas- ditionally, it classifies tumors into multiclass categories sification [ 8 ] have emerged, where they employ Convolu- (meningioma, glioma, and pituitary tumors) using a sectional Neural Network (CNN) architectures for analyzing ond dataset containing 152 MRI images collected from the medical images. Harvard repository. However, when dealing with limited

In [ 9 ], the authors introduced a method for brain tu- volumes of data, which is the case in the second dataset, mor detection using a publicly available brain tumor MRI the proposed ’23-layer CNN’ architecture faced an overfitdataset comprising data from 233 patients. In this ap- ting problem. To address this issue, they applied transfer proach, a preprocessing step was utilized to enhance the learning by combining the VGG16 architecture with the images quality, and two pre-trained deep learning models ’23-layer CNN’ architecture in a reflective manner. The were used to extract powerful features. The extracted experimental results demonstrate the efectiveness of the features were then combined into a hybrid vector using proposed models, achieving an accuracy of 97.8% . the partial least squares (PLS) method. With the aid of In another study, Aurna et al. [15] introduced an accuagglomerative clustering, the technique achieved 98.95% rate and automated brain tumor classification approach classification accuracy. using three distinct MRI datasets and a merged dataset. Ramzan et al. [ 10 ] directed their eforts toward devel- These datasets include images of three types of brain oping four sequential CNN models for classifying brain tumors (meningioma, glioma, and pituitary tumors) as tumors in MRI images. These experiments were con- well as normal brain images. The study selects the best ducted on a Kaggle dataset comprising 3,000 MRI images. models and concatenates them in two stages for feature The study involved two key steps: data preprocessing extraction. The most significant features are chosen usand automatic classification into two classes - tumor and ing Principal Component Analysis (PCA) and fed into the selected classifier. The proposed ensemble model to the MRI brain images. Subsequently, a 2D CNN and achieves an average accuracy of 99.13%. a convolutional auto-encoder network were developed Raza et al. [16] introduced a hybrid deep learning model, and trained with predetermined hyperparameters. The named DeepTumorNet, in their study for classifying three 2D CNN featured several convolutional layers, with all types of brain tumors using a basic CNN architecture. layers in this hierarchical network utilizing a 2*2 kernel The GoogLeNet architecture served as the foundation, function. Additionally, six machine-learning techniques with the last 5 layers replaced by 15 new layers in the de- were employed and compared for brain tumor classificavelopment of the hybrid DeepTumorNet approach. The tion. The obtained results indicated a training accuracy proposed model was assessed using the publicly avail- of 96.47% for the proposed 2D CNN and 95.63% for the able research dataset, known as the CE-MRI dataset. This proposed auto-encoder network. dataset consists of 3062 MRI images from 233 patients, representing three distinct types of brain tumors. The evaluation yielded an accuracy of 99.67%. 3. Proposed Methodology In [17], Deep learning architectures, including CNN, DNN, LIM (LeNet Inspired Model), AlexNet, and ResNet, In our suggested approach, we propose using a pair of were employed to classify brain MRI images as normal models to classify brain tumors into four categories: Pior abnormal. Gender and age were considered as addi- tuitary, Meningioma, Glioma, and No Tumor. The initial tional attributes to enhance the accuracy of the classifi- model utilizes a Simple CNN, while the second model cation. Multiple datasets were utilized, including those improves precision by incorporating transfer learning from Figshare, Brainweb, and Radiopaedia. The Figshare from pre-trained VGG19 and ResNet50 models. dataset comprised 1130 abnormal brain MRI images, the Brainweb dataset contained T1 weighted data with 181 3.1. Proposed Methodology of Tumor slices of normal and abnormal data, and the Radiopaedia Classification Using CNN dataset included 768 T1 images and FLAIR data. Experimental findings indicated that the LIM model demon- Convolutional Neural Networks (CNNs) are essential strated superior performance compared to SVM, AlexNet, for improving the accuracy of tumor identification in and ResNet in classifying brain MRI images as normal or medical image processing. Our goal is to develop a model abnormal with an accuracy of 82%. that accurately detects tumors from two-dimensional In addition, a segmentation and classification system brain MRI data. CNNs are preferred over fully-connected based on transfer learning is presented in [18]. It uses pre- neural networks for tumor detection due to their trained CNN (AlexNet and VGG-19) for classification, and efective parameter sharing and sparse connectivity, threshold and quick bounded box algorithms for segmen- which maximize accuracy and computational eficiency tation. The evaluation of Kaggle and Figshare datasets by utilizing features found in medical images. showed that the transferred VGG-19 and AlexNet models achieved high accuracies. Specifically, the VGG-19 model We have integrated and implemented 6 CNN layers obtained 99.75% and 98.50% accuracy, while the AlexNet specifically designed for tumor detection and classificamodel achieved 98.89% and 97.25% accuracy, respectively. tion, as illustrated in Figure 1. This combined model, These findings confirm the superior performance of the comprising 8 stages and involving the integration of VGG-19 model compared to the AlexNet model. hidden layers, has demonstrated the most remarkable Recently, Gómez-Guzmán et al.[19] utilized the Msoud outcomes in the context of tumor identification. In the dataset, which consisted of Figshare, SARTAJ, and Br35H following paragraph, we provide an overview of the datasets, totaling 7023 MRI images. The dataset com- suggested technique and a brief explanation of each of prised four classes: three brain tumor types and healthy its components. brain images. The CNN models, including Generic CNN and six pre-trained models (ResNet50, InceptionV3, In- Starting with a convolutional layer as the initial step, ceptionResNetV2, Xception, MobileNetV2, and Eficient- the input of MRI images is shaped into a uniform dimenNetB0), were trained with preprocessed MRI images us- sion of 224x224x3, ensuring consistency across all images. ing various strategies. Among all the CNN models, Incep- Once the images are standardized, a convolutional kernel tionV3 demonstrated superior performance, achieving is constructed to interact with the input layer. This keran average accuracy of 97.12% on this dataset. nel employs 32 convolutional filters, each with a size of Furthermore, Saeedi et al. [20] employed a dataset of 3x3, and operates on 3-channel tensors. The purpose is 3264 T1-weighted contrast-enhanced MRI brain images, to extract low-level features from the MRI data eficiently encompassing images of three types of brain tumors and without overparameterizing, considering the complexity healthy brains. The research commenced with the ap- and quantity of the data. We specifically use the Rectified plication of preprocessing and augmentation algorithms Linear Unit (ReLU) activation function to introduce nonour multiclass (4 classes) classification task, we opt for the softmax activation function in the final layer, as it consistently demonstrates superior accuracy compared to other options. We also employ the "categorical crossentropy" loss function. Our optimization approach of choice is "Adam", an abbreviation for "Adaptive Moment Estimation". Adam builds upon the foundations of gradient descent and integrates concepts from the Adaptive Gradient Algorithm (Adagrad) version. It adapts step sizes for each parameter dynamically during the training process using a decaying average of partial gradients.

The model underwent training for 32 epochs. However, in the current context, the accuracy falls short of our expectations. Consequently, we have decided to enhance our approach by incorporating transfer learning through ResNet50 and VGG19 architectures. The objective is to fortify the accuracy of our model and further elevate its overall performance. linearity and prevent it from aligning too closely with the output.

The ConvNet architecture undergoes a systematic reduction in spatial dimensions, efectively reducing parameter 3.2. Proposed Methodology Using count and computational load. A valuable Max Pooling transfer learning layer is adept at curbing overfitting concerns tied to brain MRI images. For geographical data that complements Within this section, we have utilized two separate models input images, MaxPooling2D is employed. A pivotal con- – VGG-19 and ResNet-50 – to tackle the complexities volutional layer operates at dimensions of 111x111x32. associated with brain tumor detection and classification. The pooling size of (2, 2) enacts vertical and horizontal We will provide detailed explanations of these two models downsizing due to input image segmentation across both in the subsequent sections. spatial dimensions.

The network comprises multiple convolutional blocks, 3.2.1. VGG-19 progressively increasing filter count to 64, 128, and 256 in The VGG network, created by Simonyan and Zisserman subsequent layers. This strategic augmentation aims to at the University of Oxford in 2014 [21], is a widely recogcapture intricate features in the input MRI images. Inter- nized pre-trained CNN model. Trained on the extensive spersed Max Pooling layers mitigate overfitting concerns. ImageNet ILSVRC dataset containing 1.3 million images The architecture concludes with a spatial dimension of divided into 1000 classes, it consists of 19 layers, includ7x7x256, signaling an abstract representation for down- ing 16 for convolutions, 3 fully connected layers, and stream tasks. This design fosters computational eficiency 5 for pooling. Instead of average pooling, Maxpooling and addresses standardization concerns for diverse MRI is used for downsampling. The fully connected layers inputs, underscoring a commitment to non-linearity and consist of two sets, each with 4096 channels, followed by prevention of overfitting in the network’s learned repre- another fully connected layer with 1000 channels for lasentations. bel prediction. Importantly, the last fully connected layer After the pooling layer, we obtain a pooled feature map. benefits from GPU acceleration during training, enabling Flattening becomes crucial at this point, as it requires the use of a softmax layer for classification. us to reshape the entire matrix that represents the input images into a single-column vector. This modification 3.2.2. ResNet-50 is necessary for further processing. We then feed this lfattened vector into the Neural Network for additional ResNet-50, an abbreviation for residual neural network, processing. is a convolutional neural network featuring a depth of In our methodology, we incorporate three dense layers: 50 layers. This model was developed and trained by He the first layer comprises 512 hidden units, the second et al. [22] in their research conducted in 2016. Similar layer consists of 256 hidden units, and the third layer to VGG-19, this model is able to classify a wide range of serves as the final layer. This sequence of 512, 256, and 4 objects, with a total of 1000 categories. The model’s trainis tailored to match the complexity of our classification ing regime capitalized on a dataset comprising more than task. To address potential overfitting risks, we introduce 1 million images sourced from the ImageNet database. a dropout rate of 50% between these hidden layers. For

4. Experimental Results

In our exploration of brain tumor detection, we extensively leveraged various brain MRI image databases to 3.2.3. Proposed CNN Architecture using Transfer construct a comprehensive dataset for the training, valiLearning dation, and testing phases of our Convolutional Neural Network (CNN) models. The dataset utilized is curated Transfer learning plays a crucial role in augmenting the from the Br35H dataset and the Chen Jung dataset [23]. base CNN model, utilizing the feature maps generated It’s important to note that the Chen Jung dataset comby the pre-trained VGG19 and ResNet50 models. Both prises three tumor classes (glioma, meningioma, pituVGG19 and ResNet50 models undergo weight fetching, itary), whereas the class without a tumor was sourced retaining the original weights acquired during the ini- from the Br35H dataset. The latter dataset originally tial training. Specifically, only the last four layers intro- contains only two classes (tumor, non-tumor). We specifduced in the subsequent training session remain train- ically extracted the non-tumor class after preprocessing able. This strategic approach ensures the preservation and analysis to seamlessly integrate it with Cheng Jun’s of pre-existing knowledge from the initial training, with dataset. This meticulous curation ensures a comprehenifne-tuning focused on the recently added layers for op- sive and diverse representation of brain MRI images for timized performance in targeted classification tasks. On our research. the flip side, we loaded pre-existing saved files from the The dataset used for training and testing our models pre-trained models (VGG19 and RESNET50 Model). Sub- consists of around 3027 T1-weighted MRI images in JPEG sequently, we concatenated these models with the pro- format.These images were thoughtfully classified into posed CNN Model to create a new model named "Con- four distinct classes: glioma, meningioma, pituitary, and catenated Model", that generates output by averaging no tumor. The following Figure 3 illustrates examples predictions from the three individual models. This ensem- from each class. ble technique is designed to improve overall prediction We partitioned the dataset into three subsets: trainperformance by capitalizing on the diverse strengths of ing, validation, and testing, with respective percentages each base model. This contributes to heightened robust- of 80%, 10%, and 10%. However, the initial image count ness and generalization capabilities across a range of data proved insuficient for efective neural network training. types. Figure 2 presents a comprehensive illustration of To address this limitation, we implemented a practical the model. solution: data augmentation. This image-processing technique enabled us to generate additional data and images from the original dataset. In the training phase, the initial count of 2419 images was augmented, resulting in a total of 4838 images. 4.2. Evaluation metrics The algorithm’s performance measures, including accuracy, were assessed using equation-defined TP (true positive), TN (true negative), FP (false positive), and FN (false negative) values.

=

+ + + + (1)

Figure 6 illustrates the training process of the ResNet model, encompassing 32 epochs, and utilizing an image size of (224x224x3). Regarding accuracy, the initial validation accuracy commenced below 69%, exhibiting a swift 4.3. Discussions and comparisons rise to nearly 77% after the inaugural epoch. Similarly, the initial validation loss surpassed 1.01 but diminished Various experimental assessments were conducted to val- to below 0.69 following the first epoch. Figure 6 visually idate the proposed dense CNN model. All experiments captures favorable accuracy enhancement and loss reducwere performed in a Python programming environment tion. Notably, the validation accuracy, initially modest, with GPU support. Initially, image preprocessing in- exhibited progressive improvement, reaching 88.7%. volved augmenting the images for training, enhancing Similarly, the VGG19 model underwent training for 32 the model’s accuracy in detecting augmented tumors. epochs, utilizing an image size of (224x224x3). ConcernThe proposed model achieved an accuracy of 97.23% on ing accuracy, the initial training accuracy commenced bethe training dataset and 97.75% on the validation dataset, low 76%, experiencing an increase to nearly 85% after the as illustrated in Figures 4 and 5. ifrst epoch. Similarly, the initial training loss exceeded

The experiments were conducted over 100 epochs, 0.74 but decreased to below 0.49 following the first epoch. with a batch size of 32 and image size set at (224x224x3). Figure 7 illustrates the positive trend of enhancing accuIn terms of accuracy, the initial validation accuracy racy and reducing loss. The validation accuracy, initially started below 61% but rapidly increased to nearly 68% modest, exhibited progressive improvement, reaching after the first epoch. Similarly, the initial validation loss almost 94.58%. was above 1.05 but decreased to below 0.80 after the first Finally, Figure 8 illustrates the positive trend in imepoch. Figure 5 depicts the positive trend in improving proving accuracy and reducing loss of the concatenated accuracy and reducing loss. The validation accuracy, model involving the three architectures: VGG19, ResNet initially low, progressively improved to almost 97.23%. 50, and a simple CNN model trained on the same dataset, The subsequent experiments involved the ResNet50 concatenated together. This model underwent training model, VGG19, and a concatenation of all three models, 03 classes encompassing the three types of brain tumors (Meningioma, Glioma, Pituitary), and the 04-class classification used in our studies, which includes the three types of brain tumors along with a class denoted as "no Tumor." Various datasets are employed for this comparative analysis. Table 2 presents the accuracy achieved by each model, with accuracies ranging from 82% to 98.78%.

Notably, all these values fall below the accuracy attained by our model, which stands at 99.34%. 5. Conclusion for 32 epochs, with a batch size of 32 and an image size set at (224x224x3). In terms of accuracy, the initial training accuracy started below 96%, experiencing an increase to nearly 97% after the first epoch. Similarly, the initial training loss exceeded 0.25 but decreased to below 0.22 following the first epoch. After completing the last epoch, the training accuracy reached 98.37%, with a validation accuracy of 99.34%.

The CNN model achieves a testing accuracy of 98.69% with a testing loss of 0.13. In comparison, the VGG19 model attains a testing accuracy of 95.76% and a testing loss of 0.10. The ResNet50 model demonstrates a testing accuracy of 96.94% with a testing loss of 0.1339.

Remarkably, when concatenating the three models (CNN, VGG19, and ResNet50), an outstanding testing accuracy of 99.34% is achieved, accompanied by a test loss of 0.17.

A detailed comparison of test accuracy and loss among diferent models is presented in Table 1

The proposed model’s accuracy is evaluated in comparison with other developed models designed for brain tumor classification. These models are assessed across three types of classifications: the binary classification of 02 classes (normal or abnormal brain), the classification of

Dataset

Br35H Dataset T1-weighted contrast-enhanced MRI images

T1-weighted Contrast-Enhanced (CE) Multiple datasets (Figshare, Brainweb, and Radiopaedia

T1-weighted dataset

Tumor Classes 03 classes 04 classes 03 classes 02 classes 04 classes

Techniques Used Accuracy closed-from metric learning algorithm (CFML) 94.68%

EficientNet CNN 98.78% 16-layer CNN 96.13%

LIM model 82% CNN + Transfer Learning (VGG19, ResNet50) 99.34%

References

[1] C. Putra , Analisis citra otak pada color-task dan word-task dalam stroop task dengan menggunakan electroencephalography (eeg), Universitas Gadjah Mada , Yogyakarta ( 2014 ).

[2] B. Pattanaik , K. Anitha , S. Rathore , P. Biswas , P. Sethy , S. Behera , Brain tumor magnetic resonance images classification based machine learning paradigms , Contemporary Oncology/Współczesna Onkologia 27 ( 2022 ).

[3] J. Seetha , S. S. Raja , Brain tumor classification using convolutional neural networks , Biomedical & Pharmacology Journal 11 ( 2018 ) 1457 .

[4] N. Kumari , S. Saxena , Review of brain tumor segmentation and classification , in: 2018 International conference on current trends towards converging technologies (ICCTCT) , IEEE, 2018 , pp. 1 - 6 . Method Cheng, Jun, et al.[ 24 ] Nayak et al. [ 13 ] Sultan et al.[ 12 ] Wahlang et al.[ 17 ] Proposed method

[5] K. N. Qodri , I. Soesanti, H. A. Nugroho , Image anal- 2022 . 08 .039. ysis for mri-based brain tumor classification using [15] N. F. Aurna , M. A. Yousuf , K. A. Taher , A. Azad, deep learning, IJITEE (International Journal of In- M. A. Moni, A classification of mri brain tumor formation Technology and Electrical Engineering) based on two stage feature level ensemble of deep 5 ( 2021 ) 21 - 28 . cnn models, Computers in biology and medicine

[6] M. Mahmud , M. S. Kaiser , A. Hussain , S. Vassanelli, 146 ( 2022 ) 105539 . Applications of deep learning and reinforcement [16] A. Raza , H. Ayub , J. A. Khan , I. Ahmad , A. S. Salama , learning to biological data , IEEE transactions on Y. I. Daradkeh , D. Javeed , A. Ur Rehman , H. Hamam, neural networks and learning systems 29 ( 2018 ) A hybrid deep learning-based approach for brain 2063-2079 . tumor classification, Electronics 11 ( 2022 ) 1146 .

[7] S. J. Pan , Q. Yang , A survey on transfer learning , [17] I. Wahlang , A. K. Maji , G. Saha, P. Chakrabarti, IEEE Transactions on knowledge and data engineer- M. Jasinski , Z. Leonowicz , E. Jasinska, Brain maging 22 ( 2009 ) 1345 - 1359 . netic resonance imaging classification using deep

[8] F. J. Díaz-Pernas , M. Martínez-Zarzuela , M. Antón- learning architectures with gender and age , Sensors Rodríguez, D. González-Ortega , A deep learning 22 ( 2022 ) 1766. approach for brain tumor classification and seg- [18] S. Gull , S. Akbar , I. A. Shoukat , A deep transfer mentation using a multiscale convolutional neural learning approach for automated detection of brain network , Healthcare 9 ( 2021 ) 153. tumor through magnetic resonance imaging , in:

[9] M. Aamir , Z. Rahman , Z. A. Dayo , W. A. Abro , 2021 International Conference on Innovative ComM. I. Uddin , I. Khan , A. S. Imran , Z. Ali , M. Ish- puting (ICIC) , IEEE, 2021 , pp. 1 - 6 . faq, Y. Guan , Z. Hu , A deep learning ap- [19] M. A. Gómez-Guzmán , L. Jiménez-Beristaín , E. E. proach for brain tumor classification using mri im- García- Guerrero , O. R. López-Bonilla , U. J. Tamayoages , Computers and Electrical Engineering 101 Perez, J. J. Esqueda-Elizondo , K. Palomino-Vizcaino , ( 2022 ) 108105 . doi:https://doi.org/10.1016/ E. Inzunza-González, Classifying brain tumors j . compeleceng . 2022 . 108105 . on magnetic resonance imaging by using convo-

[10] N. Remzan , K. Tahiry , A. Farchi , Brain tumor clas- lutional neural networks , Electronics 12 ( 2023 ) 955. sification in magnetic resonance imaging images [20] S. Saeedi , S. Rezayi , H. Keshavarz , S. R Niusing convolutional neural network ., International akan Kalhori , Mri-based brain tumor detection Journal of Electrical & Computer Engineering (2088- using convolutional deep learning methods and 8708) 12 ( 2022 ). chosen machine learning techniques , BMC Medical

[11] T. Hossain , F. S. Shishir , M. Ashraf , M. A. Al Nasim , Informatics and Decision Making 23 ( 2023 ) 1 - 17 . F. Muhammad Shah, Brain tumor detection using [21] K. Simonyan , A. Zisserman , Very deep convoluconvolutional neural network , IEEE ( 2019 ) 1-6. tional networks for large-scale image recognition,

[12] H. H. Sultan , N. M. Salem , W. Al-Atabany , Multi- arXiv preprint arXiv:1409.1556 ( 2014 ). classification of brain tumor images using deep neu- [22] K. He , X. Zhang , S. Ren, J. Sun , Deep residual learnral network , IEEE Access 7 ( 2019 ) 69215 - 69225 . ing for image recognition , in: Proceedings of the doi:10 .1109/ACCESS. 2019 .2919122. IEEE conference on computer vision and pattern

[13] D. R. Nayak , N. Padhy , P. K. Mallick , M. Zymbler, recognition, 2016 , pp. 770 - 778 . S. Kumar, Brain tumor classification using dense [23] J. Cheng, Brain tumor dataset, figshare. Dataset, eficient-net, Axioms 11 ( 2022 ) 34 . 2017 . https://doi.org/10.6084/m9.figshare. 1512427 .

[14] M. S. I. Khan , A. Rahman , T. Debnath , M. R. Karim , v5. M. K. Nasir , S. S. Band , A. Mosavi , I. Dehzangi , [24] J. Cheng, W. Yang, M. Huang , W. Huang , J. Jiang , Accurate brain tumor detection using deep con- Y. Zhou , R. Yang , J. Zhao , Y. Feng , Q. Feng , et al., volutional neural network, Computational and Retrieval of brain tumors by adaptive spatial poolStructural Biotechnology Journal 20 ( 2022 ) 4733 - ing and fisher vector representation , PloS one 11 4745 . doi:https://doi.org/10.1016/j.csbj. ( 2016 ) e0157112 .