INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue IX September 2025

www.rsisinternational.org

Page 1089

A Comprehensive Review on Brain Tumour Segmentation Using

Deep Learning Approach

Akshay, Varsha Bodade

Terna Engineering College, Neurl

DOI: https://dx.doi.org/10.51584/IJRIAS.2025.1009000106

Received: 06 October 2025; Accepted: 12 October 2025; Published: 25 October 2025

ABSTRACT

One of the most critical tasks in medical image processing is the analysis of brain tumour. Early detection of

the disease however helps and assists in enhancing prior treatments and thereby increasing the survival rates of

the patient. The process of conducting a manual segmentation through a series of MRI images of the brain

helps to diagnose cancer. It is further accompanied through clinical routines and check-ups. The procedure of

the same appears to be a tedious and time consuming job. Hence there is a need to automate the process by

using advanced technologies such as machine learning and deep learning. The fundamental aim of this paper is

to contribute in providing a detailed review of brain tumour segmentation methods along with the proposed

method for final implementation. Automated form of brain segmentation has been widely adopted by various

research scholars; wherein they use deep learning based algorithms to address the problems that remained

restricted by using ML strategies. Many algorithms and techniques that fall under DL are capable to produce

better results through objective evaluation on large amounts of MRI based patient data. With large and

extensive research being done in the domain by using ML; this review paper differs by highlighting the trend

and strategies used in DL. An introduction to brain tumours is presented in the paper followed by state-of-art

algorithms which are to be proposed during the implementation phase. An assessment on current and existing

work is also briefed and future directions to standardize the process of tumour segmentation are addressed.

Keywords: CNN, deep learning, EfficientNetB2, MRI, image processing

INTRODUCTION

A form of uncontrolled and unnatural grow of cells in the body is termed as cancer. Such occurrences leads to

the division of cells and at times; an unnatural growth in the cells are also observed. When such changes occur

in the brain tissues, the disease is called as brain tumour. While on one hand, where brain tumours are not that

common; they can at times be lethal [1]. Depending on the source of origin of the growth cells, brain tumours

can either be divided into primary and metastatic brain tumour. In primary brain tumour, the origin of the

growth cells takes place in the brain; whereas in metastatic brain tumour; the origin of the tumour could be

anywhere in the body which eventually spreads and resides into brain.

Gliomas tumour is a type of brain tumour that originates in the glial cells and appears to be as the point of

research for many scholars and authors. The word glioma is a term which is used to refer various forms and

types of tumour stages that occurs in the brain. Some of them include low intensity glioma such as

astrocytomas and oligodendrogliomas [2]. On the other hand high intensity glioma comprises of glioblastoma

multiform (GBM). Common and most observed forms of treating cancer are chemotherapy and radiotherapy

[3]. Apart from this, medical imaging techniques such as Computed Tomography (CT), Positron Emission

Tomography (PET), and Magnetic Resonance Imaging (MRI) are also used to derive valuable information

about the tumour [4]. This information is further used to understand the location, size and metabolism of the

tumour that resides within the brain.

While these clinical routines and medical imaging techniques being used; MRI is considered to be as the most

accurate and reliable technique to understand the location of the tumour which resides in the brain. MRI is a

non-invasive technique that makes use of radio frequency signals that triggers the targeted tissues and tends to

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue IX September 2025

www.rsisinternational.org

Page 1090

produce an image of those tissues under the influence of powerful magnetic fields thus generated [5]. A

sequence of brain images is generated and the modalities are later studied for tissue contrast image. Figure 1

depicts MRI modalities that represents the occurrence of gliomas in T1 weighted MRI, T2 weighted MRI and

Fluid Attenuated Inversion Recovery (FLAIR).

Figure 1: MRI Modalities showing types of glioma

The acquisition and image capturing through an MRI image can vary in different devices. An approximate

number of 120 slices are expected to occur while capturing the 3D volume of the brain image. Later, when all

the slices are eventually combined, the diagnosis of the modalities becomes complicated. In general terms, the

T1 image of the glioma is used to differentiate between the healthy and unhealthy tissues; whereas the T2

images are used to perform segmentation on the bright signal region of the edema [5]. This edema region is

responsible to detect active cell areas and further save it from tumour tissue. Finally, the FLAIR image of the

glioma modality consists of water molecules within it and is later used to differentiate between the edema

region and the region of Cerebrospinal Fluid (CSF).

Therefore it can be said that before applying any strategic therapy, it is mandatory for the doctor to check the

stages of the unhealthy tissues that tends to disrupt the normal functioning of the human brain. The

segmentation process usually involves diagnosis and separating the active tissues from unhealthy tissues; the

necrotic regions from edema regions which also include the separation of grey matter (GM) with white matter

(WM). The current process of clinical routine however comprises of segmentation, division, activation and

manual annotation of the MRI images. With clinical routines and procedures being a time consuming job, the

inclination of research scholars over robust and automated systems is highly observed.

Being in tangent with this methodology, the presented paper therefore focuses on image segmentation

techniques which are used widely. It also focuses on the deep learning algorithms which are expected to be

used during the implementation phase. A comparison of varying methods is also provided in the paper and the

review finally ends by providing a research direction for future.

LITERATURE SURVEY

This section of the paper highlights the work done by other research scholars.

Authors in [6] made use of MRI images to diagnose brain tumour. They performed the same by implementing

deep learning models such as CNN and transfer learning models. VGG16, MobileNet and InceptionV3 were

used for the same. The dataset consisted on 10,000 images with an image resolution of 200 x 200 pixels. The

dataset was further divided into two categories of healthy brain images and tumour brain images. A training

accuracy of 92percent and a testing accuracy of 93.28percent were achieved. Similar work was obtained by

authors in [7] wherein they used a DCNN model to detect brain tumours from MRI images. BRATS was used

as the dataset which comprised of 950 Gliomas images, 600 tumour images and 940 meningioma images were

present. The proposed model achieved a precision accuracy of 91.38percent by combing the model with

DCNN and VGG19. SVM was also a part of the model implementation.

Authors in [8] tried to use a cutting edge automation technique that used CNN along with a correlation

learning method (CLM). A total of 700 meningioma images and 1500 glioma images were present in the

dataset. Additionally, 3000 brain tumour images were detected in the dataset and the CLM model achieved a

precision accuracy of 91.32percent.

Authors Garg et al. in [9] proposed the implementation of ML based algorithms such as random forest, KNN,

SVM and decision trees. For this purpose a hybrid model was encouraged and ensemble classifier was used.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue IX September 2025

www.rsisinternational.org

Page 1091

The classification of brain tumours was performed on 2500 brain tumour images. Thirteen feature extraction

techniques were used and the ensemble classifier was able to generate an accuracy of 93.36percent. The

accuracy however produced by the hybrid model was observed to be 94.28percent. Authors in [10] proposed

the architecture of EfficientNetB2 along with CNN as a multi modal project. Brain tumor was diagnosed by

scanning the MRI images. The authors also made use of transfer learning based models such as ResNet50,

MobileNetV2 and InceptionV3. A total of 92percent accuracy was obtained through this model with an

F1Score of 93.68percent. Authors in [11] developed a CNN based neural network that was implemented to

diagnose brain tumours over a dataset of 2000 MRI images. BRATS was used for the same purpose and the

findings and results were quite promising. The model accomplished an accuracy of 94.26percent under 100

epochs.

Authors Khalil et al. in [12] implemented the diagnosis of brain tumour by making use of two step dragonfly

algorithm. The dataset used for the same comprised of 3D MRI images. One of the primary difficulties

observed in this research was the understanding of the dataset due to the tumour size and structure in the MRI

images. In order to overcome such challenges; a two-step method was adopted to extract a common point of

origin. BRATS 2017 dataset was used for the same with 3D images of the brain. The results and findings of

the study were further evaluated by using Fuzzy Means, SVM and decision trees as ML based algorithms.

Parameters of evaluation such as accuracy, recall and precision score were calculated. The weakness of this

study was further eliminated in a research work performed by authors in [13] wherein they used the BRATS

dataset to diagnose regions of brain tumours. The analysis and validation of the same was done in a two-phase

training method and regularization techniques such as dropout. The hybrid model thus developed attained an

accuracy of 91percent and various segmentation methods of semiautomatic approach were used. The model

was expected to perform better than CNN’s.

Authors in [14] made use of ML based techniques such as SVM and KNN to analyze abnormalities in the fetal

brain. A hybrid approach of combining RF, NB and decision trees was also proposed in the research. Once an

evaluation of the model was done, BRATS dataset was later used for training and testing purpose. An accuracy

of 91.36percetn was obtained. Authors Attallah et al. in [15] made use of various DL based algorithms to

diagnose the presence of tumour in the brain. A series of detailed architecture was used and diagnosis of

embryonic neurodevelopmental abnormalities (ENDs) was focused on.

Table 1: Comparative Analysis of various researches

Reference

Dataset

Model

Cons

GIST

[16]

11,000 MRI

images were

used

VGG-16,

MobileNet

Due to smaller

dataset,

accuracy was

not optimized

Image

augmentation is

required

[17]

3200 brain

tumour images

DCNN, SVM

Varying types of

images is

required

Augmentation

can be proposed

[18]

2500 MRI

Images

CLM + CNN

Results were not

optimized due to

imbalance in the

dataset

SMOTE could

have been used

to balance the

dataset so that

better results

could have been

obtained

[19]

2200 Images

Hybrid

ensemble

Number of

evaluation

parameters was

A pre-trained

model could

have been used

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue IX September 2025

www.rsisinternational.org

Page 1092

classifier

less. Train and

test ratios were

not split in a

decisive manner

and planned

ratios of train

test and

validation could

have been

proposed

[20]

3200 brain

tumour images

ResNet50,

MobileNet

A huge lack of

performance

metrics was

observed due to

smaller and

imbalanced

dataset

Proposal of

SMOTE and

other LDA

techniques could

have resolved

the issue

[21]

BRATS 2015

Residual

Network

Data processing

was done in a

slow manner

with not much

efficient

algorithms

which

eventually

resulted into

over-fitting of

the dataset

Over-fitting

could have been

avoided by

using LDA and

parameter tuning

techniques

[22]

BRATS 2019

DCNN, VGG19,

ENB2

Individual

tumours per

slice were

trained and

tested which

increased the

overall training

and testing time

of the model

Time

complexity

could have been

reduced by

ensuring

planned ratio

split of the

dataset

Methods adopted for Image Segmentation

The methods used for segmenting brain tumour can be classified into three broad categories: manual, semi-

automatic and fully automatic [23]. However the methods to be used in a specific project are determined by the

level of interaction the project demands. Depending on that a specific category of brain tumour segmentation

method can be selected.

3.1 Manual Method: this method is generally adopted by radiologists to extract tumour information from MRI

images and further perform multi-modality segmentation on them through anatomical and physiological

knowledge. However, the process to do the same involves the radiologist to clinically examine the tumour on

MRI image, slice by slice and then diagnose the tumour and segregate the edema region. This method not only

appears to be a time consuming task; but rather also involves dependency on the radiologist to examine the

same.

3.2 Semi-Automatic Method: this method is generally adopted to initialize, intervene and generate feedback

from the evaluation process thus done [24]. The process of initialization is done after segregating the tumour

region from the healthy regions of the brain. It is in this stage, that various pre-processing methods are applied

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue IX September 2025

www.rsisinternational.org

Page 1093

to the tumour dataset. Additionally, processing algorithms such as LDA and PDA are also used in this stage so

that the project can be steered towards a desired result. The generated results are then evaluated and altered if

not satisfactory. Once this process of initialization is done, a cellular automata (CA) is implemented on the

segmented images so that a probabilistic vector mapping can be done. This approach of adapting to vector

maps and later generating tumour images is called as MRI Modality through FLAIR. Results from

initialization and CA are then combined to determine a final volume of the tumour [25]. In this stage, various

DL and ML based classifiers are used so that data can be trained and tested on MRI images. The results

however generated contain within them a series of noise and redundant data and must be removed so that

tissue types from the brain image can be enhanced. For this purpose, special voxels from the dataset are

selected and spatial coordinates are formed. The voxels thus created are in alignment to the corresponding type

of the tissue. SVM’s and CNN’s are majorly used in this stage.

3.3 Fully Automatic Method: this type of brain segmentation method does not require any user interaction to

be present in order to detect the tumour. However, there are certain challenges associated with the method. The

primary challenge is to determine the gliomas in the MRI image. This segmentation of gliomas is a difficult

task since the size and location of the gliomas tumour would vary from patient to patient. Also, with no

boundaries and irregular regions in the MRI image, it becomes difficult to scan the edged in the MRI image.

Additionally, introducing the manual efforts from radiologists would stretch the time and dramatically expose

the health condition of the patient, if not treated on time. Hence, in such a scenario a fully automatic

segmentation method is used which is performed by slicing different sections of the image and adjusting the

intensities of every slice so that several modalities of the sliced image can be obtained and worked

individually. The fully automatic segmentation method is further classified into descriptive and generative

methods [26].

Descriptive: the descriptive method is a classification technique that tries to establish a relationship between

the data input and its associated instances. A series of feature extraction and pre-processing stages are applied

Generative: the generative method on the other hand, works on probabilistic models who tends to differentiate

between healthy tissues an unhealthy tissues by locating their spatial coordinates and determining the region of

disease

METHODOLOGY

This section of the paper briefs on the expected methodology which is to be implemented. It also summarizes

on a workflow diagram and takes into account the usage of DL models.

Proposed Architecture

The primary aim of the expected research is to classify the occurrence of various types of brain tumours by

using a GUI platform. The dataset consisting of MRI images is initially obtained from the Kaggle repository.

In the next step, a series of data pre-processing is applied such as labeling the data, resizing the shape of the

image etc. Once the dataset is pre-processed, a validation system is applied on the dataset wherein the data is

divided and split into a planned ratio for training, testing and validation stages. The next step involves

application of DL and transfer learning based algorithms such as CNN, EfficientNetB2 and VGG16. On

application of algorithms, the programming model is able to classify the input given as one of the four types of

brain tumour. In the end, results are calculated on the basis of accuracy and recall factors and finally,

performance analysis of the same is determined. The entire classification application takes place on GUI

platform using Flask; wherein the user can enter the image and the model shall classify it as either of one of

the four types of brain tumour. Figure below explains the workflow of the research.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue IX September 2025

www.rsisinternational.org

Page 1094

Figure 2: Proposed Workflow

Environmental Set up

The expected methodology is proposed to be performed on Flask wherein the user can upload an MRI image

and the model can further classify it as either healthy image of the brain or presence of tumour in the brain.

The platform makes use of 32GB RAM and the system is expected to train deep learning algorithms and

generate optimized and efficient results.

Dataset Used

The dataset required to implement the research is obtained from Kaggle repository [27]. MRI images are

present in the dataset. A total of 920 Meningioma (a), 240 no tumour (b), 500 pituitary tumor (c) and 800

glioma (d) tumor MRI images are expected to be used in the training and testing phase of the model. Figure

below displays the overview of the dataset.

Figure 3: MRI Dataset

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue IX September 2025

www.rsisinternational.org

Page 1095

Data Pre-Processing

Pre-processing of data is one of the most essential and critical stages of any detection model. Once the dataset

from the repository is obtained, it undergoes a few stages of pre-processing such as resizing the image,

labeling the image, balancing the data etc. This is done so that high quality of database can be used and the

MRI Images can be normalized so that efficient result generation can take place. Additionally, in order to

smooth the MRI images obtained from the repository, the blurred mages from the background are removed so

that a sense of image sharpness can be obtained in the original image. Various techniques such as Gaussian

and Laplace filters are used to adopt the same.

Data Division

In scenarios where only small amounts of dataset is used; attaining better results and optimization becomes a

challenging task. Since the working of neural networks requires large size of dataset to induce enhanced

results, techniques such as augmentation and data division are used. Augmentation is a process of generating

more images of data form a single source of dataset which is obtained through rotational practices. When

images are horizontally, vertically and diagonally flipped, rotated, zoomed or shifted; a series of images are

produced. This process is called as augmentation. However, the method can be fruitful if less number of MRI

images is obtained from the repository. On the other hand, the process o f division leads to a planned and

strategized categorization of dataset into ratios of training, testing and validation. Both the techniques can

however be used to enhance the learning capacity of the model.

Deep Learning Models

Figure 4: CNN Architecture

CNN: The research study begins by using an input image from the dataset with a pixel size of 32 x 32. This

image is sent to a convolutional layer with 16 filters. The image is then expected to feature map the MRI

images with a resolution of 32 x 32 x 16. A kernel size of 3 x 3 vector is used and the output is then forwarded

to the next layer of the CNN which appears to be the max pooling layer. The pixels from the MRI Image are

feature mapped and filter values of 32 is applied to the same. The repetition of these layers takes place in

repetition wherein the convolutional layer and the pooling layer are feature mapped continuously so that a

newly created 3 dimensional array vector could be created. The final layer is responsible to generate the output

by making use of the activation function such as Softmax. Once the last layer generates the output, a dropout

function of 0.5 is triggered and the CNN’s architecture is configured. The figure below depicts the architecture

of CNN:

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue IX September 2025

www.rsisinternational.org

Page 1096

VGG16: this algorithm appears to belong to transfer learning based model proposed by authors in [28]. A pre-

trained model and heavy structure of the network is used to diagnose brain tumour, hence the usage of VGG 16

is justified by the author. It is capable to detect and classify heavily loaded datasets and can categorize the

images with better and greater accuracy. One of the primary reasons of using VGG16, is that it performs better

while dealing with larger datasets and attains optimization with complex data. A total of 5 layers are observed

in the algorithm wherein the image of pixel resolution of 32 x 32 is used. ReLu is adapted to be as the

activation function in this model and is further used to activate the hidden layers. Figure below depicts the

architecture of VGG16.

Figure 5: VGG16 Architecture

CONCLUSIONS

Detecting the presence of tumour in the brain at the right stage plays a vital role in preventing the rates of

mortality. With varying sizes, location and structure of the tumour, the process of detection becomes a

challenging task and is therefore influenced by multiple factors. In the early stages, the process of detection is

done manually by doctors and technicians. However in complex scenarios, identification and segmentation of

the tumour is a tedious task. Therefore analyzing, recognizing and categorizing become mandatory. To address

issues of the same, algorithms and techniques of deep learning and transfer learning are used. A CNN based

classification model for brain tumour detection is therefore proposed wherein the results are obtained by using

MRI images obtained from Kaggle repository. The expected limitation of the research is the restriction of

hidden layers and GPU capacity. Future work can be performed to better correctly identify brain cancers by

using individual patient information gathered from any source.

REFERENCES

1. De Angelis L M. Brain Tumors. N. Engl. J. Med. 2001; 344:114-23

2. Deimling A. Gliomas. Recent Results in Cancer Research vol 171. Berlin: Springer; 2009.

3. Stupp R. Malignant glioma: ESMO clinical recommendations for diagnosis, treatment and follow-up.

Ann Oncol 2007; 18(Suppl 2):69-70

4. Drevelegas A and Papanikolou N. Imaging modalities in brain tumors Imaging of Brain Tumors with

Histological Correlations. Berlin: Springer; 2011; chapter 2:13-34

5. Menze B, et al. The Multimodal brain tumor image segmentation benchmark (brats). IEEE Trans Med

Imaging 2015; 34(10):1993-2024

6. Almadhoun, H.R.; Abu-Naser, S.S. Detection of Brain Tumor Using Deep Learning. Int. J. Acad. Eng.

Res. (IJAER) 2022, 6, 29–47

7. Musallam, A.S.; Sherif, A.S.; Hussein, M.K. A New Convolutional Neural Network Architecture for

Automatic Detection of Brain Tumors in Magnetic Resonance Imaging Images. IEEE Access 2022, 10,

2775–2782

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue IX September 2025

www.rsisinternational.org

Page 1097

8. Woźniak, M.; Siłka, J.; Wieczorek, M. Deep Neural Network Correlation Learning Mechanism for CT

Brain Tumor Detection. Neural Comput. Appl. 2021, 1–16. Available

online: https://link.springer.com/article/10.1007/s00521-021-05841-x

9. Amin, J.; Sharif, M.; Haldorai, A.; Yasmin, M.; Nayak, R.S. Brain tumor detection and classification

using machine learning: A comprehensive survey. Complex Intell. Syst. 2022, 8, 3161–3183

10. Nayak, D.R.; Padhy, N.; Mallick, P.K.; Zymbler, M.; Kumar, S. Brain Tumor Classification Using

Dense Efficient-Net. Axioms 2022, 11, 34

11. Obeidavi, M.R.; Maghooli, K. Tumor Detection in Brain MRI using Residual Convolutional Neural

Networks. In Proceedings of the 2022 IEEE International Conference on Machine Vision and Image

Processing (MVIP), Ahvaz, Iran, 23–24 February 2022; pp. 1–5

12. Khalil, H.A.; Darwish, S.; Ibrahim, Y.M.; Hassan, O.F. 3D-MRI brain tumor detection model using

modified version of level set segmentation based on dragonfly algorithm. Symmetry 2020, 12, 1256.

13. Sajid, S.; Hussain, S.; Sarwar, A. Brain tumor detection and segmentation in MR images using deep

learning. Arab. J. Sci. Eng. 2019, 44, 9249–9261

14. Lotlikar, V.S.; Satpute, N.; Gupta, A. Brain Tumor Detection Using Machine Learning and Deep

Learning: A Review. Curr. Med. Imaging 2022, 18, 604–622

15. Attallah, O.; Sharkas, M.A.; Gadelkarim, H. Deep learning techniques for automatic detection of

embryonic neurodevelopmental disorders. Diagnostics 2020, 10, 27

16. Urban G. et al. Multi-modal brain tumor segmentation using deep convolutional neural networks.

MICCAI Multimodal Brain Tumor Segmentation Challenge (BraTS) 2014:31–35

17. Zikic D. et al. Segmentation of brain tumor tissues with convolutional neural networks. MICCAI

Multimodal Brain Tumor Segmentation Challenge (BraTS) 2014:36–39

18. Havaei M, Davy A, Farley W D, Biard A, Courville A, Bengio Y, Pal C, Jadoin P M, Larochelle H.

Brain tumor segmentation with deep neural networks. Medical Image Analysis 2016

19. Davy A. et al. Brain tumor segmentation with deep neural networks. MICCAI Multimodal Brain

Tumor Segmentation Challenge (BraTS) 2014:1–5

20. Pereira S, Pinto A, Alves V, Silva C A. Brain tumor segmentation using convolutional neural networks

in MRI images. IEEE Trans Med Imaging 2016;35(5):1240–1251

21. Dvorak P, Menze B. Structured prediction with convolutional neural networks for multimodal brain

tumor segmentation. MICCAI Multimodal Brain Tumor Segmentation Challenge (BraTS) 2015:13–24.

22. Rao V, Sarabi M S, Jaiswal A. Brain tumor segmentation with deep learning. MICCAI Multimodal

Brain Tumor Segmentation Challenge (BraTS) 2015:56–59

23. Gordillo N, Montseny E, Sobrevilla P. State of the art survey on MRI brain tumor segmentation. Magn

Reson Imaging 2013; 31(8):1426–38

24. Foo JL. A survey of user interaction and automation in medical image segmentation methods. Tech rep

ISUHCI20062, Human Computer Interaction Department, Iowa State Univ; 2006

25. Havaei M, Larochelle H, Poulin P, Jadoin P M. Within-brain classification for brain tumor

segmentation. Int J Cars 2016; 11:777-788

26. Bauer S, Wiest R, Nolte L, Reyes M. A survey of MRI-based medical image analysis for brain tumor

studies. Phys Med Biol.2013;58:97-129

27. https://www.kaggle.com/datasets/sartajbhuvaji/brain-tumor-classification-mri

28. Simonyan, K.; Zisserman, A. Very deep convolutional networks for large-scale image recognition.

arXiv 2014, arXiv:1409.1556