UMT Artificial Intelligence Review (UMT-AIR)

Volume 2 Issue 2, Fall 2022

ISSN(P): 2791-1276 ISSN(E): 2791-1268
Homepage: https://journals.umt.edu.pk/index.php/UMT-AIR

Article QR

Multiclass Light Weight Brain Tumor Classification and Detection

Title:
using Machine Learning Model Yolo 5
Asif Raza1, Usman Amjad2, Muhammad Abubakr3, Dr. Humera4, Asad Abbasi5,
Author (s):
Asher Ali1
1
Syed University of Engineering and Technology, Karachi, Pakistan
Affiliation (s): 2
NED University of Engineering and Technology, Karachi, Pakistan
3
MNS-University of Agriculture, Multan, Pakistan.
4
University of Karachi, Pakistan
5
Benazir Bhutto Shaheed University Lyari, Karachi, Pakistan
DOI: https://doi.org/10.32350.umt-air.22.04

History: Received: September 15, 2022, Revised: November 23, 2022, Accepted: December 15,
2022
A, Raza, U. Amjad, M. Abubakr, Humera, A. Abbasi, and A. Ali, “Multiclass
Citation: light weight brain tumor classification and detection using machine
learning model Yolo 5,” UMT Artif. Intell. Rev., vol. 2, no. 2, pp. 00–00, 2022,
doi: https://doi.org/10.32350.umt-air.22.04

Copyright: © The Authors

Licensing: This article is open access and is distributed under the terms of
Creative Commons Attribution 4.0 International License
Conflict of
Interest: Author(s) declared no conflict of interest

A publication of
Department of Information System, Dr. Hasan Murad School of Management
University of Management and Technology, Lahore, Pakistan
 Multiclass Lightweight Brain Tumor Classification and Detection
using Machine Learning Model Yolo v5
Asif Raza1*, Usman Amjad2, Muhammad Abubakr3, Dr. Humera4, Dr.
Asad Abbasi5, Asher Ali1
Department of Computer Science and Information technology, Sir Syed
1

University of Engineering and Technology, Pakistan

2
Department of Computer Science and Information Technology NED University
of Engineering and Technology, Pakistan
3
Department of Computer Science, MNS-University of Agriculture, Multan,
Pakistan
4
Department of Computer Science, University of Karachi, Pakistan
5
Benazir Bhutto Shaheed University Lyari, Karachi, Pakistan
Abstract- Early brain tumor accuracy. For this purpose, data
identification is a critical challenge augmentation was applied to the
for neurologists and radiologists. publically available dataset from
Manually identifying brain tumors Kaggle. MRI of different classes
through magnetic resonance imaging including 396 glioma images, 397
(MRI) is difficult and prone to meningioma, 380 no tumor, and 399
mistakes. The diagnosis of tumor is a images of pituitary tumors were
complex job when performed in a employed. The current study
traditional manner. Brain presents false negative, true positive
abnormalities can be fatal, lowering false positive, and true negative,
a patient's quality of life and which were used to test the YOLO v5
adversely harming their overall (You Only Look Once) classifier
health. Brain tumors vary in nature performance. It was determined that
based on where they are situated and the YOLO v5 model is giving 88%
how rapidly they develop inside the accuracy.
skull. Tumors are a proliferation of
Index Terms- machine learning,
abnormal nerve cells that form a
mass. Some brain tumors begin in magnetic resonance imaging (MRI),
the cells that support the brain's medical imaging, single shot
nerve cells. This paper proposes a detection (SSD), YOLO v5.
machine learning algorithm known
as YOLO v5 SSD (single shot
detection) to detect and classify such
tumors namely meningioma, glioma,
and pituitary gland with 88%

*
Corresponding Author: [email protected]
 I. Introduction having bounding and labeled boxes
possessing the same characteristics.
Brain tumors comprise the
It's significantly lighter and quicker.
formation of abnormal tissues inside
In the past, this model was trained
human brain. Brain tumors can be
on the “COCO” dataset. Its structure
classified through MRI into
consists of 24 convolutional layers
different types based on their
for feature extraction from images
mutation, location, and genetic cell
and two fully connected layers for
composition. The most common
bounding and labeled boxes
primary brain tumors are
prediction by using the dark net
meningioma, glioma, and pituitary
framework.
tumors [1], as shown in Figure 1.
Medical imaging is an important
application of computer vision,
where accurate segmentation of
actual lesion symptoms is crucial for
diagnosis and treatment. One
effective technique for achieving
accurate segmentation is the Grab
Cut method, which is capable of
segmenting out specific areas of
interest. Moreover, the fine-
tuning method Introduces to obtain
Fig. 1. Brain Tumor MRI image [2]
image aspects along with the hand-
Brain illnesses, such as tumors, crafted
are a primary cause of mortality (color, size) features. Feature
and disability because they damage optimization was achieved using
the principal cells in the human entropy for accurate and rapid
brain. Brain abnormalities increase classification. The proposed model
the risk of brain tumors which are was verified using databases from
the ninth leading cause of death prominent medical images and
across the world. computing.MICCAI dataset was us
YOLO v5 was used to construct ed along with BRATS for
the object detection model. This segmentation and Analysis [4].
model is managed by employing the This study presents a unique
dark net framework. It offers a metaheuristic-based technique for
single network used for tumor detection. The proposed
classification and prediction by technique contains classification,
segmentation, and feature extraction approximately 212 MRI samples
by utilizing a deep neural network. [7].
Furthermore, this method is also MRI analysis is yet again a
compared with the existing matrix common approach for detecting
e.g. correct detection rate (CDR) an brain tumors. These images are
evaluation matrix, false acceptance trained of a new hybrid approach
rate (FAR), and False Rejection rate that combines Neural
(FRR) indices, It was determined Autoregressive Distribution
that the recommended strategy Estimation (NADE) with CNN,
outperforms the other existing which was used to test 3064 T1-
algorithms [5]. weighted contrast-enhanced scans
"Inception-v3" is associated of three different classes. The
with "DensNet201". Two results revealed that when medical
alternative scenarios of brain tumor images are rarely available, then
classification and detection were hybrid CNN-NADE can be
examined using these two models. employed with outstanding
The softmax classifier was further classification performance [8].
applied to categorize features. Brain tumors are difficult to
Secondly, features were extracted detect. Recent literature suggests
from multiple "DensNet" segments that more work needs to be done to
using the pre-trained "DensNet improve detection rates. MRI scans,
201[6]. which are commonly used to detect
This research provides several brain tumors, are prone to acquiring
Segmentation techniques e.g. Guide noise during image acquisition.
Segmentation and Semi-automatic Removing this noise is a complex
Segmentation are introduced for task that requires careful attention to
cerebrum tumor diagnosis by using detail. Despite advancements in
CAD. Using a GLCM matrix, imaging technology, there is still a
texture-based features were need for further research and
retrieved. Correlation and development to improve the
homogeneity are some of the accuracy of brain tumor detection.
textural features of the images With continued effort and
considered in particular. Multi-layer innovation, it is hoped that future
perceptron (MLP) and Naive Bayes methods will be more effective in
algorithms were combined to get the detecting brain tumors and
highest accuracy of 98.6% and improving patient outcomes [9]–
91.6% respectively by analyzing [11].
 Segmentation is an important to detect the orientation of the fetal
technique used in medical research brain is particularly noteworthy, as
to gain the full understanding of the this information can be crucial for
structure and behavior of tumors. accurate diagnosis and treatment
The current study focused on brain planning. The use of a machine
tumors and used YOLOv5, a state- learning algorithm to evaluate the
of-the-art object detection, detection and classification of
classification, and segmentation abnormalities, such as
model. Four classes of images malformation, is also a promising
including normal brain images, as approach. Machine learning
well as brain images with algorithms can be trained on large
meningioma, glioma, and pituitary datasets of labeled images to learn
tumors were incorporated in this patterns and features that are
study. The main objective was to indicative of specific abnormalities.
classify, detect, and segment the Then, they can be used to accurately
tumors more accurately. The results identify these abnormalities in new
showed that YOLOv5 was able to images [12].
effectively segment the tumors in
This study offers a method for
brain images. In fact, the
automatically recognizing and
performance of YOLOv5 was faster
segmenting brain tumors. Seven
and more accurate than the previous
light weight versions of Yolo are
versions of YOLO.
used for detections are
This paper is organized as segmentation. The use of neural
follows: Section I comprises networks for detection and
Introduction, Section II presents the segmentation is a common approach
related work, Section III contains in computer vision. This article
experimental results, and Section IV proposed, seven different neural
presents the conclusion and future networks perform effective
directions. segmentation task. Each of the
seven neural networks likely has a
II. Related Work
specific architecture and is
Precise epic localization optimized for a particular aspect of
algorithm in the YOLO v4 the detection and segmentation
architecture for fetal brain MRI process. The use of multiple
analysis represents a promising algorithms provides a
approach for detecting and comprehensive evaluation of the
classifying healthy and abnormal effectiveness of different
fetal brains. The algorithm's ability approaches, which can help to
identify the most promising method outcomes. The proposed model's
for their further development and superior performance in detecting
refinement. The use of popular and brain tumors from MRI images
widely used frameworks, such as demonstrates its potential for
YOLO v3, YOLO v4, Scaled improving medical image analysis,
YOLO v4, YOLO v4 Tiny, YOLO while aiding in more precise
v5, Faster-RCNN ensured that the diagnosis and treatment planning
results remained relevant and for patients with brain tumors.
comparable to other studies in the Compared to other CNN-based
field. After training and evaluating models, such as AFP-Net, Mask
the models on 641 MRI scan images RCNN, YOLOv5, and FCNN, the
from the dataset, the YOLO v5 proposed model demonstrated better
model was determined to provide performance in detecting brain
the best performance. This suggests tumors.
that the YOLO v5 algorithm may be Deep learning architecture was
particularly effective in brain tumor developed for brain tumor
detection and segmentation. classification using MRI images.
Furthermore, it could be a The framework consists of three
promising area for future research main stages: data set pre-processing,
and development [13]. deep learning feature extraction, and
Overall, the proposed algorithm classification. The data set pre-
holds a significant potential for processing stage involves image
processing large sets of brain tumor normalization, resizing, and data
images and providing quick and augmentation to improve the quality
accurate outcomes for medical and quantity of the dataset. The deep
diagnosis and treatment planning. learning feature extraction stage
The faster convergence rate and the uses the VGG19 model as feature
higher accuracy achieved by the extractor, followed by either a CNN,
proposed model makes it a GRU, or Bi-GRU model for further
promising approach for improving feature extraction. The
the efficiency and effectiveness of classification stage employs the
brain tumor detection and extracted features to classify brain
classification [14] tumors into three types. The three
models used for classification are
Given the potentially life-
VGG19 + CNN, VGG19 + GRU,
threatening nature of brain tumors,
and VGG19 + Bi-GRU [16].
their accurate and efficient detection
is crucial for improving patient The U-Net architecture has been
widely used and has achieved state- approach. Furthermore, It employs
of-the-art results in brain MRI ensemble learning for enhanced
segmentation tasks. Several efficiency, resulting in the highest
improved versions of U-Net have possible accuracy in tumor detection
been proposed to further improve [19]. In this paper, a method for
the segmentation performance. For tumor segmentation using FLAIR
this purpose, data comprising lower- MRI and FCNNs is suggested. In a
grade glioma collections with previous study, FCNN produced
minimal FLAIR can be particularly tumor-containing sub-regions in the
useful, as this is a challenging task original picture and provided
due to the subtle nature of these segmented full-size FLAIR to help
tumors. This paper proposes a better radiologists to enhance their
U-Net Architecture for brain tumors diagnosis [20].
using VGG16. In this regard, it The ability to accurately
upgraded the U-Net Architecture distinguish between firm and soft
with VGG-16 by using k-fold cross
meningiomas is critical in
validation on the TCGA-LGG
determining the appropriate
dataset to segment MRI images and
treatment strategy for patients. The
identify tumor cell regions [17].
proposed deep learning approach
Tumor detection and that uses YOLO v4 can aid in
segmentation using MRI images improving the accuracy of
remains difficult and error-prone. diagnosis, leading to more
As a result, for the early appropriate patient counseling and
identification of the disease, a tumor operative procedures. Furthermore,
detection system is necessary. This the superior performance of the
study proposes two deep learning YOLO v4 model over the traditional
methods for tumor classification classification methods, such as
diagnosis using the cutting-edge SVM and RF, highlights the
object framework and the DL potential of deep learning-based
library FastAi. The BRATS 2018 approaches for a more accurate and
dataset was used in particular, efficient medical image analysis
which includes 1,992 MRI scans [21].
[18]. Based on an encoder-decoder
DCNN model for brain tumor architecture, a semantic
segmentation differs from the segmentation network for tumor
existing CNN models in the sense sub-region segmentation from 3D
that it employs a trial-and-error MRIs was established. A variational
auto-encoder branch was included tasks, making it a valuable tool for
to recreate the input images, owing various medical industries.
to a restricted training dataset size,
IV. Results and Findings
in order to regularize the shared
decoder and to place extra A typical strategy in machine
limitations on its layers [22]. The learning uses 80% of the images for
current study proposes feature training and 20% for validation.
recombination for semantic This split is designed to train the
segmentation which uses linear model on several images, while still
expansion and compression to build reserving a group of images separate
more complex features, as well as for validation to make absolutely
a segmentation SE (SegSE) block sure that the result is conclusive to
for feature recalibration that avoid over fitting. Having 4
captures contextual data while different classes, this particular
preserving spatial meaning. model shows a relatively better
Furthermore, it evaluates the average accuracy of around 88.4%.
proposed approaches for brain A. YOLO
tumor segmentation using publicly
available data [23]. The current object identification
model was built with YOLO v5. It
III. Methodology was managed by the dark net
YOLO v5 is a newer version of framework that offers a single
YOLO, designed for segmentation. network for both item classification
This model provides highly and prediction through bounding
competitive results due to being boxes. This particular version is
lightweight in nature. The YOLO now substantially faster and lighter.
algorithm takes a dataset in the form This model was rapidly trained on
of pictures and corresponding text the unique annotated MRI images. It
files. The dataset is divided into featured fully connected layers for
different folders for training and bounding box prediction. This
validation. The .yaml extension file network was built more efficiently
is used to define paths and classes using the dark net framework [24].
and is mandatory in the newer This model offers particular benefits
YOLO versions. After proper in its design.
processing and segmentation, i. Comprehensive object
results can be obtained. YOLO v5 is identification, tumor placement,
a highly efficient model that can as well as fast detection speed
accurately perform segmentation and accuracy.
ii. Recognizing microscopic tumor system are true positives (TP),
objects in murky, noisy, and false positives (FP), true negatives
hazy pictures. (TN), and false negatives (FN).
The following matrices stand to
B. Use of Simulation Software
calculate the performance. The
In the current study, Google- ability to accurately distinguish
Colab was used to run and train between different forms of brain
classification techniques. The tumors is measured by accuracy
accuracy rate was initially quite low [25].
for all models, as shown in Figure 5.
To evaluate a test's accuracy,
Although, it rose as epochs
increased up to 142. Epoch was the percentage of true positive and
executed with a batch size 16. true negative occurrences in all
examined instances was
The current study used a determined using the formula given
confusion matrix to classify the below. A true positive (TP) arises
relationship between distributions when the model accurately predicts
and data in order to examine the the positive class. An outcome
performance of the classification where the model properly predicted
process. The precision rate was the positive class is referred to as a
above 88%. Classification may be true positive. Similar to a True
evaluated in a broad spectrum by positive, a true negative is a result
analyzing different confusion for which the model accurately
matrices. The study includes four expects the negative class. When the
basic keys: “true positive”, “true model forecasts the positive class
negative”, “false positive”, and inaccurately, it is called a false
“false negative”. Model positive as shown in Figure 2 shows
performance was computed in terms the confusion matrix Yolo v5
of specificity, accuracy, sensitivity, model. Confusion matrix visualizes
negative predictive precision, and summarizes the performance of
values, and the F1-scores classification algorithm. It shows
Mentioned in Figures 3, 4, 5, and 6, the relevance of various classes, for
respectively. instance, 0.42% accuracy for
C. Performance Metrics glioma, 0.92% accuracy for
meningioma, 0.77% accuracy for
The four major parameters used pituitary cancer, and 0.99%
to evaluate the efficacy of the accuracy for no tumor class.
Fig. 2. Confusion Matrix of 4 Classes
Figure 3 shows the accuracy of classification algorithms in
a model tested on Kaggle situations where classes are heavily
(augmented dataset) based on F1. imbalanced. Similar to ROC curves,
Figure 4 displays the number of precision-recall curves provide a
positive results, correctly classified graphical representation of the
out of the total number of tests that classifier's performance over
conclude as positive. It is important multiple thresholds, rather than a
to note that a few wrongly classified single value. Finally, in Figure 6, the
positive results can decrease the number of positive tests accurately
accuracy slightly. In Figure 5, the classified as positive is compared to
precision/recall curve is used to the total number of positive tests.
show the performance of parallel

Fig. 3. F1 Score
Fig. 4. Precision

Fig. 5. Precision

Fig. 6. Recall
 Figure 7 describes the number of the prediction of validation batch,
images provided into the dataset of while Fig 10 contains labels,
brain tumors, containing 4 classes in instances, and image intensity
the graph format. Fig 8 describes the coordinates levels. Fig 11 shows the
pair plot and Labels Correlogram of results and Fig 12 represents
the provided dataset. Figure 9 shows Training Batch 1.

Fig. 7. Number of Images

Fig. 8. Labels Correlogram

Fig. 9. Validation Batch Prediction

Fig. 10. Labels

Yolo v5 performs well in terms standard in machine learning. It is
of validation loss and precision good to see that the algorithm meets
accuracy. A validation loss of less this standard.
than 0.5 is considered a good
Figure 11. Results
The precision accuracy of more bounding boxes are detected
than 88.41% indicates that the sufficiently during the validation
algorithm makes accurate process. Overall, these results are a
predictions. The efficacy of YOLO testament to the effectiveness of the
v5 model is also evident in the algorithm and the YOLO v5 model
validation and prediction batch being used.
shown in Figure 9, where the

Fig. 12. Training Batch 1

 D. Comparison of Results • Metric / Precision Accuracy
The results of YOLO v5 model with • Metric / Recall Accuracy
other existing results were
Considering the above
compared on the basis of the
parameters, the model performed
following parameters:
well in terms of the reduction of
• Training loss training loss, validation per object
loss, metric/precision accuracy, and
• Validation per object loss
training / class loss.
• Training Loss / Class
Table I
Please provide caption here
Validation Training
Training
Ref /Object / Class Metric/Precision Metric/Recall Tool
loss
Loss loss
0.13-
[26] 0.004 0.15 87 % 0.8 NVIDIA
0.04
0.10- Tesla/Google
[12] 0.0050 0.01 84.30 % 0.8
0.01 Colab
0.10-
Proposed 0.012 0.01 88.41 0.6 Google Colab
0.01

V. Conclusion and Future Work a larger dataset with different size

and resolutions to improve
The current study proposed a
algorithm accuracy.
method for the detection and
classification of MRI in 4 different References
classes. Tumor progression, [1] Johns Hopkins Medicine.
position, and region were extracted “Brain tumor types.” Johns
by using YOLO v5 SSD algorithm. Hopkins Medicine.
Tumor class such as glioma, https://www.hopkinsmedicine.o
meningioma, pituitary tumor, and rg/health/conditions-and-
no tumor were determined from the diseases/brain-tumor/brain-
dataset. The study yielded relatively tumor-types
better accuracy with this algorithm.
The proposed work is divided into [2] National Cancer Institute.
two phases, namely (i) classification “Diffuse midline gliomas
and (ii) feature extraction. diagnosis and treatment.” NCI.
Furthermore, the aim is to develop a https://www.cancer.gov/rare-
new method in the future along with brain-spine-
 tumor/tumors/diffuse-midline- https://doi.org/10.1109/ACCES
gliomas S.2020.2978629
[3] Z. Krawczyk and J. Starzynski, [7] P. Kshirsagar, A. N. Rakhonde,
“Bones detection in the pelvic and P. Chippalkatti, “MRI
area on the basis of YOLO image based brain tumor
neural network,” in 19th Int. detection using machine
Conf. Comput. Prob. Elect. learning,” vol. 81, pp. 4431–
Eng., Banska Stiavnica, Sep. 4434, Dec. 2019.
2018, pp. 1–4. doi: [8] R. Hashemzehi, S. J. S.
https://doi.org/10.1109/CPEE.2 Mahdavi, M. Kheirabadi, and S.
018.8506970 R. Kamel, "Detection of brain
[4] T. Saba, A. S. Mohamed, M. El- tumors from MRI images base
Affendi, J. Amin, and M. Sharif, on deep learning using hybrid
"Brain tumor detection using model CNN and
fusion of hand crafted and deep NADE," Biocybernet. Biomed.
learning features," Cog. Syst. Engi., vol. 40, no.3, pp. 1225–
Res., vol. 59, pp. 221–230, Jan. 1232, Sep. 2020, doi:
2020, doi: https://doi.org/10.1016/j.bbe.20
https://doi.org/10.1016/j.cogsys 20.06.001
.2019.09.007 [9] M. A. Khan, et al., "Brain tumor
[5] A. Hu and N. Razmjooy, “Brain detection and classification: A
tumor diagnosis based on framework of marker‐based
metaheuristics and deep watershed algorithm and
learning,” Int. J. Imaging. Syst. multilevel priority features
Technol., vol. 31, no. 2, pp. 657– selection," Microscop. Res.
669, Jun. 2021, doi: Tech., vol. 82, no. 6, pp. 909–
https://doi.org.10.1002/ima.224 922, Feb. 2019, doi:
95 https://doi.org/10.1002/jemt.23
238
[6] N. Noreen, S. Palaniappan, A.
Qayyum, I. Ahmad, M. Imran, [10] U. Agrawal, E. N. Brown,
and M. Shoaib, “A Deep and L. D. Lewis, “Model-based
learning model based on physiological noise removal in
concatenation approach for the fast fMRI,” NeuroImage, vol.
diagnosis of brain tumor,” IEEE 205, Art. no. 116231, Jan. 2020,
Access, vol. 8, pp. 55135– doi:
55144, 2020, doi: https://doi.org/10.1016/j.neuroi
 mage.2019.116231 resonance imaging,” Int. J. Elec.
Comput. Eng., vol. 13, no. 1, pp.
[11] Y. K. Dubey and M. M.
1039–1047, Feb. 2023, doi:
Mushrif, “FCM clustering
https://doi.org/10.11591/ijece.v
algorithms for segmentation of
13i1
brain MR images,” vol. 2016,
Art. no. 3406406, doi: [16] A. M. G. Allah, A. M.
https://doi.org/10.1155/2016/34 Sarhan, and N. M. Elshennawy,
06406 “Classification of brain MRI
tumor images based on deep
[12] S. Paul, D. T. Ahad, and M.
learning PGGAN
Hasan, “Brain cancer
augmentation,” Diagnostics
segmentation using YOLOv5
(Basel), vol. 11, no. 12, Art. no.
deep neural network,” arxiv,
2343, Dec. 2021, doi:
doi:
https://doi.org/10.3390/diagnost
https://doi.org/10.48550/arXiv.221 ics11122343
2.13599
[17] S. Ghosh, A. Chaki, and K.
[13] N. M. Dipu, S. A. Shohan, Santosh, "Improved U-Net
and K. M. A. Salam, “Brain architecture with VGG-16 for
tumor detection using various brain tumor
deep learning algorithms,” in segmentation," Phys. Eng. Sci.
2021 Int. Conf. Sci. Contem. Med., vol. 44, pp. 703–712 May
Technol., Dhaka, Bangladesh, 2021, doi:
Aug. 2021, pp. 1–6. doi: https://doi.org/10.1007/s13246-
https://doi.org.10.1109/ICSCT5 021-01019-w
3883.2021.9642649
[18] N. M. Dipu, S. A. Shohan,
[14] S. Arunachalam and G. and K. M. A. Salam, “Deep
Sethumathavan, “An effective learning based brain tumor
tumor detection in MR brain detection and classification,” in
images based on deep CNN 2021 Int. Conf. Intell. Technol.,
approach: i- YOLOV5,” Appl. Hubli, India, Jun. 2021, pp. 1–6.
Artif. Intell., vol. 36, no. 1, Art. doi:
no. 2151180, Dec. 2022, doi: https://doi.org/10.1109/CONIT
https://doi.org/10.1080/088395 51480.2021.9498384
14.2022.2151180
[19] M. B. Naceur, R. Saouli, M.
[15] A.-A. Nayan, et al., “A deep Akil, and R. Kachouri, “Fully
learning approach for brain automatic brain tumor
tumor detection using magnetic
 segmentation using end-to-end Cham, Springer, 2019, doi:
incremental deep neural https://doi.org/10.1007/978-3-
networks in MRI images,” 030-11726-9_28
Comput. Meth. Prog. Biomed., [23] S. Pereira, A. Pinto, J.
vol. 166, pp. 39–49, Nov. 2018, Amorim, A. Ribeiro, V. Alves,
doi: and C. A. Silva, “Adaptive
https://doi.org/10.1016/j.cmpb. feature recombination and
2018.09.007 recalibration for semantic
[20] P. R. Lorenzo et al., segmentation with fully
“Segmenting brain tumors from convolutional networks,” IEEE
FLAIR MRI using fully Trans. Med. Imaging, vol. 38,
convolutional neural networks,” no. 12, pp. 2914–2925, Dec.
Comput. Meth. Prog. Biomed., 2019, doi:
vol. 176, pp. 135–148, Jul. 2019, https://doi.org/10.1109/TMI.20
doi: 19.2918096
https://doi.org/10.1016/j.cmpb. [24] M. Hammami, D. Friboulet,
2019.05.006 and R. Kechichian, “Cycle
[21] N. F. Alhussainan, B. B. GAN-Based data augmentation
Youssef, and M. M. Ben Ismail, for multi-organ detection in CT
“A deep learning approach for Images Via Yolo,” in 2020
brain tumor firmness detection IEEE Int. Conf. Image Proc.,
using YOLOv4,” in 2022 45th Abu Dhabi, United Arab
Int. Conf. Telecommun. Signal Emirates, Oct. 2020, pp. 390–
Proc., Prague, Czech Republic, 393. doi:
Jul. 2022, pp. 342–348. doi: https://doi.org/10.1109/ICIP407
https://doi.org/10.1109/TSP556 78.2020.9191127
81.2022.9851237 [25] K. Muhammad, S. Khan, J.
[22] A. Myronenko, “3D MRI D. Ser, and V. H. C. de
brain tumor segmentation using Albuquerque, “Deep learning
autoencoder regularization," in for multigrade brain tumor
Brainlesion: Glioma, Multiple classification in smart
Sclerosis, Stroke and Traumatic healthcare systems: A
Brain Injuries. BrainLes 2018. prospective survey,” IEEE
Lecture Notes in Computer Trans. Neural Netw. Learning
Science, A. Crimi, S. Bakas, H. Syst., vol. 32, no. 2, pp. 507–
Kuijf, F. Keyvan, M. T. Reyes, 522, Feb. 2021, doi:
and T. van Walsum, Eds., https://doi.org/10.1109/TNNLS
 .2020.2995800 with fine-tuning approach,”
Comput. Mathemat. Meth. Med.,
[26] T. Shelatkar, Urvashi, M.
vol. 2022, pp. 1–9, Jul. 2022,
Shorfuzzaman, A. Alsufyani,
doi:
and K. Lakshmanna, “Diagnosis
https://doi.org/10.1155/2022/28
of brain tumor using light
58845
weight deep learning model