Deep Learning-Based Methods for Brain Tumor Segmentation: A State-

of-the-Art Review

Syed shahid Abbas 1, Salahuddin2*, Abdul Manan razzaq3, Mubashar hussain4, Meiraj
aslam5, Prince Hamza shafique6 and Muhammad Asif Nadeem 7

1,2,3,,5,6
Department of Computer Science, NFC Institute of Engineering and technology, Multan, Pakistan.
4
Department of computer science, university of Engineering and technology Lahore Pakistan.
7
Department of English, Institute of Southern Punjab, Multan,Punjab.

*
Corresponding Author: Salahuddin. Email: [email protected]

_____________________________________________________________________________

Abstract:

Hospitals have lately begun using machine learning to expedite the diagnostic
and analysis process. Now that they have assistance with diagnosis, doctors can
expedite the start of the healing process. AI in healthcare may be used for simple to
complex tasks in the future, such as phone answering, reviewing medical records,
trending and analytics in primary care, computer design and therapeutic medicine,
reading radiology images, creating treatment and diagnosis plans, and even having
conversations with patients. Medical imaging such as CT, MRI, and X-ray pictures may
be interpreted using deep learning models to establish a diagnosis. Inconsistencies and
dangers can be identified by the algorithms in the medical imaging data. Cancer
detection frequently makes use of deep learning. Brain tumors must be correctly
segmented using MRI images in order to aid in clinical diagnosis and therapy planning.
However, the lack of certain diagnostic procedures in MRI images makes medical
practice more challenging. The recommended method performs better when comparing
the quantitative and qualitative results of medical image analysis as it is currently
performed. When it comes to the accurate identification of malignant lung nodules in the
event of lung cancer detection, CT scans of the chest perform better. Early detection of
lung cancer is crucial for patients' chances of survival. Using sparse chest computed
tomography (CT) data from earlier research, create a multi-view knowledge-based
collaborative (MV-KBC) deep model to distinguish between benign and malignant
nodules. However, the MV-KBC model had more accuracy. Nevertheless, the model
can only be used to supervise image data. In this research, we present a novel deep
learning-based multi view model to alleviate the model's shortcoming. The accuracy of
the suggested model was significantly improved, and computation and classification
times were decreased, for semi-supervised medical image applications.

Keywords: classification, MV-KBC, inconsistencies, computed tomography (CT),

algorithms

1. Introduction
Over the years, interest in deep learning methods for image analysis has
increased. One specific area is healthcare. Medical imaging involves the use of 3D
images of the human body, usually obtained from CT or MRI [1] scanners. Doctors
investigate and discover the patient's condition based on medical records. Thus, they
can diagnose

medical or surgical procedures. Using deep learning to analyze medical images

is a revolutionary process [2]. Using deep learning for high-quality medical image
analysis. Many deep learning algorithms can be used to analyze data obtained from
diagnosis. Deep learning is used to learn how to process medical images using
clustering, which includes detection, segmentation, and classification. In medical image
analysis, three different techniques are often recognized: image reconstruction, image
aggregation, and bracketing. The constant use of medical image analysis [3] for raw
data can lead to errors from one task to another, especially when the data is
inconsistent. The use of medical imaging in medical decision making is increasing, but
using imaging as a pipeline often overlooks the level of image optimization. This is an
important step because successful segmentation and other subprocesses depend on
good clinical images. Digital photography is essential for everyday use. The term “image
processing” refers to how a computer processes images [5]. This work includes many
processes and procedures such as image capture, storage and presentation. A program
that shows how to measure the nature of a scene, such as color or light. Digital images
have many advantages: fast and cheap processing, easy transfer and storage, instant
measurement, multiple copies without quality intervention, faster and cheaper and
flexible exchange. Image processing requires high performance and data storage.
Despite quality control, digital images can sometimes be violated [6]. It is important to
know that research using “3D Medical Multimodal Segmentation Networks” has
significant limitations. First, the diversity and accessibility of medical records are
important for the effectiveness of these networks. There is not enough good information
from different sources and patients hinder the network’s ability to be broad. Also, the
quality and consistency of human descriptions have a great impact on accurate
segmentation, leading to the possibility of inconsistency and error, especially for difficult
things. The diversity of deep learning models can make them computationally
demanding, which can limit their applicability and accessibility in resource-limited
environments. Interpretation remains a challenge, as these standards often come down
to consensus, which can cause problems in important clinical applications where
understanding the logic behind the decision is important. The risk of overfitting and the
ability of models to generalize to new imaging or unpublished data illustrate how difficult
it is to achieve robust and reliable segmentation. The use of such models is complicated
by the need to resolve ethical and regulatory issues and provide good clinical data for
translating scientific research into clinical practice. Finally, given the rapid pace of
development of deep learning, it is important to keep up with new models and methods
to make research valid and relevant. Edit photos on your computer. This method has
the advantages of communication, data storage, flexibility, and adaptability. The
development of image scaling technology to improve photo editing. This method
requires simultaneous analysis of images on multiple systems. Both 2D and 3D images
can be made in multiple dimensions [7]. The first image processing technology was born
in the 1960s. This technology has many applications, such as image enhancement in
television, medicine, space exploration, and art. In the 1970s, when computers first
appeared, images were made faster and cheaper. In 2000, imaging became better,
easier and faster [8]. This method uses tracking techniques, transfer learning, data
augmentation and the development of neural network topologies. This is an important
task in the analysis of medical images, the identification of brain tumors, treatment
planning and follow-up, because of the accuracy of the classification. Deep learning has
great potential in this area because it can extract important features from data without
the need for special engineering. Because clinical models vary in imaging, data
acquisition, and clinical content, there are many unique issues and concerns when
training cross-sectional models to evaluate different medical procedures, including X-
rays, MRI, and CT scans.

Multi-detector row CT (MDCT) is widely used to identify and confirm lung disease
[9]. It is considered the gold standard for diagnosing these diseases and is particularly
sensitive in diagnosing lung nodules. MRI can identify 85% to 95% of nodules
measuring 5 to 11 mm [10]. Depending on the risk of lung cancer, caution is advised for
lesions larger than 7 or 8 mm, even if MDCT shows only 1 or 2 mm in size. Bacteria
smaller than 7 mm in diameter should be present to determine the growth pattern [11].
Koyam et al. [16] stated that non-contrast lung MRI (Figure 1) is a better method for
detecting malignant tumors than thin-section MDCT. There was no significant difference
between the total diagnosis and the diagnosis of malignant nodules (p>0.05). 97.0%,
p0.05, will give lower numbers than MDCT.
 Fig 1. Normal and Affected patient’s images

The main purpose of the "MV-KBC deep model" is to determine the positive and
negative aspects of chest CT images. Determining whether nodules, especially cancer,
are benign (non-cancerous) or malignant (cancerous) is an important task to improve
the standard nodule classification accuracy by combining a model with multiple clinical
quality, anatomical data utilization and aggregation of features from multiple images.
The use of anonymous data may be particularly useful in cases where recorded data
are difficult and expensive to obtain. Treatment decisions are not affected by the
presence or absence of lymph nodes in any ipsilateral bronchial region or hilum (which
may indicate N1 disease). Ipsilateral mediastinal or subcarinal lymphadenopathy is a
manifestation of N2 disease and occurs when only one region is affected. Stage of
classification from N1 to N3 indicating severity of disease. N3 describes the main stage
of disease. In N3 disease consisting of Virchow, scalene or hilar lymph nodes, major
surgery is not recommended [12]. Printers and images combine MRI and PET [13].
However, significant advances have made it possible to make patterns and physical
applications using single-detection images. The functional information provided by MRI
does not imply an additional dimension compared to dynamic MDCT. Although MRI
cannot measure glucose metabolism, it is considered equivalent to functional and
molecular analysis (14).
 Fig.2 Different stages of Brain tumor

The evolution of cancer is shown in Figure 2. Machine learning techniques are used to
detect cancer early to benefit patients. Brain tumors usually take months or years to
develop. Glioblastoma is the most common and deadly form of brain cancer. It is one of
our main models because its development is not detected by the immune system. It
causes fatigue and weakness, a constant need for sleep and a tendency to stay in bed
or rest for days. Weight loss and loss of muscle mass. Difficulty swallowing food or
drinks and little or no appetite. Artificial intelligence (AI) is driving change and growth in
many areas of healthcare. MRI, CT scan and X-ray. AI can share a wealth of patient
information, enabling early diagnosis and treatment. AI is also using automation to
develop new drugs, predict disease outbreaks, and streamline operations. AI-powered
telemedicine and remote patient care are facilitating virtual consultations and
information gathering. Genomic analysis helps identify genetic factors related to health
issues, while natural language processing (NLP) improves communication and
electronic medical records. As AI integration continues to transform healthcare, careful
attention to ethics, privacy, and regulatory issues is required, suggesting collaboration
between AI experts, professionals, and physicians. In order to ensure patient safety,
ethical concerns, and equity in new treatments, standards and standards for responsible
and equitable use of image analysis should be established. Magnetic resonance
imaging (MRI) signals can identify tumors [15]. According to a recent study, lung cancer
can be identified by DWI and is different from the findings detected by DWI after
obstructive lobe collapse [17]. Whole-body MRI combined with DWI can be used to
detect M stage in cancer patients with an accuracy comparable to PET-CT. In addition,
quantitative DWI analysis can be used to distinguish infected from non-infected tumors
[18].
2. Literature
In today's environment, brain tumors are more dangerous than cancer. Early detection
is important for screening and treatment planning. MRI scanning is the best technique
for brain tumors because it creates similar tissue without the need for radiation therapy.
In terms of analysis, traditional brain tumor segmentation methods are quite effective.
This study uses machine learning techniques and capabilities. However, brain
segmentation is still difficult, especially when there are not many methods. There are
three different groups of missing patterns. The first category refers to differences in the
brain structure of the patients [19]. The second problem is that gliomas vary in size,
shape and texture from patient to patient. Differentiating between different low-intensity
magnetic resonance imaging techniques. H. Fu et al., “Brain segmentation with missing
nodes in MRI” [20]. This approach overcomes the problem of segmentation
incompleteness with the help of robust segmentation and latent representation learning.
However, many segmentation methods make successful fusion difficult [21]. Anatomical
landmarks and landmark identification are prerequisites for the interpretation of medical
images. In this process, various methods are used to diagnose patients and plan their
care. This process includes measurement analysis, segmentation method configuration,
and image registration. Although seemingly simple, manual identification of anatomical
locations is sometimes difficult and time-consuming. Automatic localization methods are
faster and more accurate than manual identification, and are especially useful when
there are a large number of image locations that need to be processed accurately.
Latent feature extraction is a recently introduced segmentation method for resolving
missing nodes. The problem is where to get to the core of the work and how to learn it.
The HeMIS implementation [22], using Havaei’s real-time framework, independently
examines the appearance of each sample and then calculates the fixed features across
multiple changes to estimate the final segmentation. However, analyzing the mean and
variance of each representative does not help you discover the hidden representatives.
The mean function is used to determine the separation. Chen et al. distilled the ideas
into numerical expressions and mathematical concepts. Uncertainty operations are
used in [24]. Then, through the gating operation, the content code is put into a shared
representation for segmentation. Although the two coders need to match, this method is
more difficult and time-consuming.
Accurate and timely identification of anatomical structures in examination and
diagnosis is essential for the treatment of patients. The two biggest problems of these
systems are the use of poor-quality engineering methods and the low performance of
human search algorithms. J. Liu [26] proposed a system that uses artificial intelligence
to detect viruses. Deep learning principles are also used in this field to help obtain
images as input to estimate human anatomy. The model will be developed to
characterize various image analyses. In addition, many morphological objects in
scanned images are also used to distinguish target anatomical objects in our body.
However, the deep learning method performs better than the state transition method
[27]. However, it is important to increase the accuracy. Then, the model will calculate
the relationship between them to improve the performance analysis idea. The deep
network in the experiment uses features between recurrent and convolutional neural
networks for classification. Tests showed that single-view mode increases efficiency
and effectiveness in thinking about cancer. It shows great potential in Covid-19. It
learned to use multiple views of a chest CT image. Deep learning 28 is currently the
best image recognition method, but it requires multiple training models, which is not
usually used in clinical settings. The authors of this study used a small amount of chest
computed tomography data to distinguish breast cancer from breast cancer using
multimodality integration (TMME). The technology uses the ResNet-50 model to send
image data, and then uses the Image Net database to predict lung nodes. Apply weight
adjustment to nodes when backpropagation errors are performed in the distribution.
Image, voxel value and appearance are among the metrics produced by these three
models. . The classification of lung nodules and tissues around cancer is important.
Therefore, it is very difficult to use machine learning correctly. To solve this problem, the
authors proposed a multilayer learning model based on 3D convolutional neural network
(MMEL-3DCNN). The number of patients worldwide increased by 26% in 2017.
Classification of lung nodules before diagnosis is important, especially since
computerized classification can help doctors to reach an agreement. CT image
classification can be done quickly and accurately using modern machine learning and
computer algorithms. Low-dose chest computed tomography (LDCT) is expected to
reduce the risk of death in people with early-stage disease. The accuracy and efficiency
of cancer diagnosis will be increased with AI management that is equal to or better than
human experts in terms of analytical ability. Technology: We developed an artificial
intelligence (AI) framework for cancer diagnosis using deep neural network (DNN) [29].
First, a semi-automatic annotation method is used to map the image. Then, a DNN-
based malicious classification and pulmonary nodule (LN) identification model is
developed to identify cancer, especially lung cancer, using LDCT images. The LN
discovery was called “deep learning” and was validated on large datasets using deep
learning models. Tense core muscles differently. Negative nonlinearities exist between
two models or systems. Reddy et al. [30], for example, use deep learning to determine
which images have the most accuracy. In some cases, such as when generating CT
images from MRI data, we use synthetic data to create libraries containing the original
images. This can be useful since there is no electricity. We collected a large number of
MRI images with different tumors, locations, shapes and appearances to be able to
accurately represent the model. We continue to work on the SVM classifier and various
optimization methods (softmax, RMSProp, sigmoid, etc.) to analyze our work. Our
solution was built using TensorFlow and Keras because Python is a fast programming
language.

ANNs with multiple hidden input and output layers are called deep neural
networks (DNN). Based on these techniques and methods, we can construct a hidden
layer of deep neural networks (DNN). DNN architecture provides a compositional model
that represents objects as hierarchical compositions of primitives. More importantly,
complex data can be modeled with fewer units compared to shallow networks.
Additional layers make the combination lower. Generally speaking, DNNs are
feedforward networks, meaning that data flows directly from input to output or from layer
to layer. Convolutional Neural Network (CNN) CNN is a neural network with a design
consisting of many layers. Symptoms include hormonal changes, blood clots,
weakness, unsteady gait, slurred speech, mood swings, and blurred vision. The location
of the tumor determines its type, and timely diagnosis can prolong the patient's life [31].
Benign tumors are tumors that cannot invade neighboring tissues. They can be
completely eliminated and are unlikely to return. Even if brain tumors do not spread to
other tissues, they can cause serious, permanent brain damage and death. A malignant
brain tumor is a bad place. They divide rapidly and spread throughout the brain or spinal
cord, multiplying in some places. MRI scans use powerful magnets and high-frequency
radio waves to provide accurate information about tissues. An X-ray beam is used to
perform a tomography scan. Image preprocessing, feature extraction, segmentation,
and postprocessing are the processes involved in identifying brain disease.
2. Technology
 i. CNN
There are two parts of CNN: pooling layer and convolution layer. Both of these
are considered as components of convolutional neural network. It is important to use
research to develop models and implement them. The design will suggest the structure
of CNN using many neurons. A useful strategy to learn how to design neural networks is
to study successful implementations [33]. This is possible because CNN has been
intensively researched and applied in the ImageNet Large-Scale Visual Recognition
Challenge (ILSVRC) from 2012 to 2016. Rapid progress in innovation. Since we deal
with the small size of each brain slice, a simple and effective cascaded convolutional
neural network (C-ConvNet/C-CNN) is proposed in the next step. This C-CNN model
uses two different methods for local and global objects. A special remote intelligent
tracking (DWA) mechanism is also proposed to improve the accuracy of tumor cells
over other models. The DWA method describes the central location of the brain in the
structure and effect of the tumor [35].

Fig 3. CNN architecture for image processing

Fig 3 explains about the CNN architecture for image processing. It contains the
various layers which will segment and filter the image frames.

ii. RNN
Random neural network (RNN) is represented by a network algorithm that uses
continuous data or time series [36]. Deep learning concepts are used in applications
such as Siri, voice search, and Google Translate. This effective device is also used
in medical treatment to protect human life [37]. In this model, researchers first create
a model with the required features. The feed forward strategy in RNN is used to exit
from its memory, which will affect the current input and output. A series of layers are
used to process the results collected from the previous set. Therefore, the
relationship and potential benefits are treated as a coefficient value. Many hidden
levels are where these values can be changed. The RNN output is always based on
the early stage of the network. However, due to the many layers that combine the
results, the prediction of the image value is accurate.
 Fig 4. The RNN layers structure

iii. BIDIRECTIONAL RECURRENT NEURAL NETWORKS (BRNN):

This alternate RNN network structure is described in Fig. 4 [38]. Bidirectional

RNNs take prediction performance into account to increase the accuracy of predictions
made by unidirectional RNNs, which might only use historical data. Returning to the
previous example of "feeling under the weather," if the model had predicted that
"weather" would be the final word in the string, it would have been more successful in
predicting "under".

iv. Long short-term memory (LSTM):

This famous RNN design was developed by Sepp Hochreiter and Juergen Schmidhuber
as a gradient descent solution. In other words, if the previous state influences the current
prediction, the RNN model will not be able to predict the current state correctly. It is not
suitable for peanuts. The specific nature of the nut allergy will determine whether nuts
are present or not. However, if the data comes from many previous columns, the RNN
will have difficulty integrating the input. The network's ability to predict the output
depends on these gates controlling the flow. For example, if a gender of speech such as

"he" occurs multiple times in the previous sentence, you can remove it from the hand
state.

v. Gated recurrent units (GRUs):

This RNN version is similar to LSTMs in that it makes an effort to address the
problem with temporary memory that RNN models have. Instead of "cell states," it uses
hidden states to control information, and it has two gates instead of three: a reset gate
and an update gate. Like the gates in LSTMs, the reset and update gates control how
much and what kind of information to keep.

vi. GNN

Generative adversarial networks (GANs) are an interesting new advancement in

machine learning [40]. GANs, or generative models, create new data instances based
on your training data. For example, GANs are able to generate photographs that seem
like real-world snapshots, even when the faces in those images aren't actually those of
any real individuals. Generative techniques are widely used in picture synthesis.
Autoregressive models, variational autoencoders, and GAN are examples of the most
recent generation of models [41]. A GAN's strength is its ability to produce sharp
images, but its training process is unreliable. The three most popular GAN techniques
were tested in order to synthesize nodular pictures. Initially, the genuinewGAN-GP (26)
and pix-to-pix (27) deep convolutional models were employed. Nevertheless, these
artificially manufactured pictures were fuzzy and of poor quality. Through the use of
sliced Wasserstein distance loss, international segment, and pixel normalisation, we
developed a growing more gradually wGAN inspired by pgGAN. It stabilized the training
process and generated high-quality images.

Fig 5. Image pooling in GAN

 The discriminator also had nine convolution layers. Starting with a 64-pixel
image, it went through four blocks consisting of two convolution layers before producing
the 44-pixel feature map [42] that is seen in Figure 5. With the generator, a mirror image
of this operation was created. Mini-batch discrimination was added to the convolution
network's last layer in order to improve training stability. Two compact, perfectly linked
layers were applied to this feature map in order to determine the final true or false
target.

3: Comparison Analysis:

The multi-view model is compared with single-shot imaging and traditional imaging
methods. The same data were used to ensure consistency, and the performance of
each model was evaluated using the previous method. They also performed different
levels of noise simulation and data augmentation to test the performance of the model.
segmentation. The network combines information from various measurement methods
to improve the accuracy of segmentation tasks. It is designed to process three-
dimensional medical data such as CT scans or MRI volumes. Features and models
are included. Leveraging multimodal data, which can include multiple imaging
viewpoints such as T1-weighted, T2-weighted, and FLAIR MRI scans, allows the
network to collect additional information that increases the accuracy and reliability of
the segmentation results. Such connectivity is important for identifying objects of
interest, including organs, tumors, or lesions. Combining data from multiple sources
makes it easier for the network to resolve issues arising from differences in image
texture and appearance, resulting in more accurate and reliable segmentation.
Network models can often include encoder-decoder models with cross-connections or
tracking systems. Pre- and transfer learning models can also be used to improve
network performance, especially when training data is scarce. Metrics such as
Hausdorff distance, Jaccard index, and dice coefficient, which measure the overlap
between predicted and true segmentations, are often used to evaluate the
performance of the network. They play an important role in disease assessment and
treatment procedures. However, to achieve good results in certain medical tasks, good
training on a large number of data and careful hyperparameter tuning are required for
successful deployment.

4: Proposed System

Diagram of Latent Correlation Acquiring Knowledge for Brain Tumor Partitioning

with Absent MRI Features. Magnetic resonance imaging (MRI) is a common tool used in
the evaluation of brain tumors. Accurately classifying brain tumors using magnetic
resonance imaging is essential for both clinical diagnosis and therapeutic planning.
Further information for accurate brain tumor segmentation may also be obtained using
multi-modal MR imaging [43]. It is common in clinical practice to ignore many imaging
modalities. In this work, we provide a novel method for segmenting brain tumours with
missing modalities. Since there is a strong correlation across several modalities, a
correlation model is recommended to explicitly represent the latent multi-source
correlation. If a modality is absent due to the learned correlation representation, the
segmentation becomes more dependable. First, the unique representation that every
encoder generates is used to estimate the modality independent parameter. Next, every
single representation is transformed by the correlation model into a latent representation
of a multi-source correlation.

Fig.6 Brain segmentation

Figure 6 illustrates how brain tumors may be statistically analyzed to better

understand their characteristics and develop treatment plans. Accurate lesion
segmentation necessitates the use of multiple imaging modalities with different
contrast levels. Therefore, for bigger studies, manual segmentation—
unquestionably the most accurate segmentation technique—would be unfeasible.
 Deep learning's unprecedented efficacy has made it a feasible alternative for
quantitative analysis in recent times. Nonetheless, there are unique challenges with
medical image analysis.

i. Fusion Strategy

In this paper, we examine a set of deep multimodal fusion methods in the context of
point guidance. In other words, given more video inputs (such as thickness and color
information), our goal is to process information from multiple streams when describing
motions. We specifically focus on the construction of shared representations across
layers, which has been neglected in previous studies. In the past, late fusion was used
to combine joint fusions for each model. We investigated the late fusion technique and
used the C3D [44] backbone, which has shown good results in investigating various
aspects. In addition, we evaluate the convergence of the middle levels of the network
and provide a simple way to join the early execution by analyzing 111 convolutions of
different processes. Finally, we present a new model called C3Dstitch that can learn to
combine the signals of two neurons at any level using parallel lines. Dynamic methods
use late fusion (a method often used in action recognition) to combine the results of two
or more networks at the end of the connection process. Three different training models
were investigated: 1) final integration of multiple networks, calculating the loss after
summing the average; - the method combining progress and fine-tuning. The models
used for the backbone were trained separately with the same training as the network for
each change, except for the positive stage, which was trained in several stages [46].
The main goal of this research is the technology of exchanging information on special
maps of the middle layer of the network. Our first impulse is to separate the flow at the
lowest level and then combine it into a later model. Simple convolutions combined with
multiple output extraction features form a simple fusion concept. Each input of two
combined modules must be the same as the input of the shared module in the lower
layer (post-fusion). Therefore, we cut half of the output filter in each of the 111
convolutional layers (i.e., subtract the number of flows of the filter). Convolution can be
used to reduce the dimensionality of the filter space. As a result, the final architecture
consists of three parts: 1) a general network that uses the combined representation in
the final stage, 2) two early networks for each variable.

ii. Fusion on multiple Levels via Cross-stitch Units

The model phase where the streams would be combined up until this point had to
be explicitly chosen. Our objective in this paragraph is to develop a paradigm that
permits simultaneous knowledge exchange on several levels without limiting the
locations of individual or group learning. We provide a novel multi-stream methodology
where each modality has its own unique C3D network that interacts with the others at
the fully connected and pooling layers. In this design, the output of every layer is
combined using a learned weighted average called cross-stitch units [47]. Put
differently, all networks interact pairwise at every level, and the degree of interaction
between foreign modalities is found all along the way.

Fig 7. Cross stitch unit in network layers

iii. 3D Medical Multi-modal Segmentation Network Guided by Multi-source

Correlation Constraint:
To improve the segmentation results, the relationship between several
modalities may be taken into account in the field of multifunctional segmentation
[48]. In this work, we propose a subclassification scheme with a correlation
constraint. Our network contains N model-independent encoding paths with N image
inputs, feature fusion, decoding, and correlation constraint. The model independent
encoding approach may be used to record the modality-specific properties of the N
modalities. Since there is a strong correlation between the various modalities, we
first insist on using a linear correlation block to determine the correlation between
them. Next, we use a loss function to aid the network in learning the related features,
using the linear correlation block as a guide.
 This block instructs the network to identify the latent associated properties that
are more valuable for segmentation. Since not all of the features collected from the
encoders are suitable for segmentation, we propose to change the features along
the modalities and geographical routes utilizing a dual concentration suggested
fusion block. This can assist in highlighting relevant information and hiding less
helpful information. In the end, the segmentation result is generated by projecting
the fused feature representation through the decoder [49].

Fig 8. Correlation constraints

5: Conclusion

Early detection of brain tumors is important for cancer diagnosis, because

accurate diagnosis can ensure survival. Three-dimensional (3D) magnetic resonance
imaging (MRI) is frequently used in deep learning-based brain cancer research due to
the problems associated with tumor formation. The model-free coding method can be
used to record the modality-specific behavior of N modalities. Since there is a
relationship between variables, we first report the relationship between variables to
examine their relationship. Then, it instructs the network and uses the loss function to
primarily understand the characteristics of the connection. The network is guided by the
relevant blocks to determine the most useful features for segmentation. Since not
everything received by the encoder can be used to identify the separation, we
recommend using two face-based fusion blocks to restore features relative to the path
and body. This can help highlight important information and hide less useful information.
The following are research gaps that need to be filled in the future. A special deep
learning-based brain segmentation technique is used on MRI images of brain tumors to
diagnose the disease. We are developing deep learning to reduce computational and
time allocation in semi-supervised medical images and improve detection accuracy. The
process of integrating all MRI systems simultaneously has the greatest potential, is
beneficial for future medical research and can help doctors diagnose Cancer.
Reference
1. Razzak, M.I., Naz, S. and Zaib, A., 2018. Deep learning for medical image
processing: Overview, challenges and the future. Classification in BioApps,
pp.323-350.
2. Shen, Y. and Gao, M., 2019, June. Brain tumor segmentation on MRI with
missing modalities. In International Conference on Information Processing in
Medical Imaging (pp. 417-428). Springer, Cham.
3. Zhang, D., Huang, G., Zhang, Q., Han, J., Han, J. and Yu, Y., 2021. Cross-
modality deep feature learning for brain tumor segmentation. Pattern
Recognition, 110, p.107562.
4. Chattopadhyay, A., & Maitra, M. (2022). MRI- based Brain Tumor Image
Detection Using CNN based Deep Learning Method. Neuroscience Informatics,
100060.
5. Li, L., Zimmer, V. A., Schnabel, J. A., & Zhuang, X. (2022). Medical image
analysis on left atrial LGE MRI for atrial fibrillation studies: A review. Medical
Image Analysis, 102360.
6. Byra, M., Han, A., Boehringer, A. S., Zhang, Y. N., O'Brien Jr, W. D., Erdman
Jr, J. W., ... & Andre, M. (2022). Liver fat assessment in multiview sonography
using transfer learning with convolutional neural networks. Journal of Ultrasound
in Medicine, 41(1), 175-184.
7. Chen, X., Wang, X., Zhang, K., Fung, K. M., Thai, T. C., Moore, K., ... &Qiu, Y.
(2022). Recent advances and clinical applications of deep learning in medical
image analysis. Medical Image Analysis, 102444.
8. Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., ...
& Xu, D. (2022). Unetr: Transformers for 3d medical image segmentation. In
Proceedings of the IEEE/CVF Winter Conference on Applications of Computer
Vision (pp. 574-584).
9. Jin, Y., Zhao, Q., & Cheng, J. (2022, February). Oral Cone Beam Computed
Tomography Images Segmentation Based On Multi-view Fusion. In 2022
7th International Conference on Intelligent Information Technology (pp. 59-64).
10. Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H. R., & Xu, D. (2022).
Swinunetr: Swin transformers for semantic segmentation of brain tumors in mri
images. In International MICCAI Brainlesion Workshop (pp. 272-284). Springer,
Cham.
11. Soomro, T. A., Zheng, L., Afifi, A. J., Ali, A., Soomro, S., Yin, M., & Gao, J.
(2022). Image Segmentation for MR Brain Tumor Detection Using Machine
Learning: A Review. IEEE Reviews in Biomedical Engineering.
12. Vasudevan, A., & Preethi, N. (2022). Analysis of MRI Images to Discover Brain
Tumor Detection Using CNN and VGG-16. In International Conference on
Ubiquitous Computing and Intelligent Information Systems (pp. 415-422).
Springer, Singapore.
13. [13] Li, X., Ma, S., Tang, J., & Guo, F. (2022). TranSiam: Fusing Multimodal
Visual Features Using Transformer for Medical Image Segmentation. arXiv
preprint arXiv:2204.12185.
14. Musallam, A. S., Sherif, A. S., & Hussein, M. K. (2022). A New Convolutional
Neural Network Architecture for Automatic Detection of Brain Tumors in
Magnetic Resonance Imaging Images. IEEE Access, 10, 2775-2782.
15. Shah, H. A., Saeed, F., Yun, S., Park, J. H., Paul, A., & Kang, J. M. (2022). A
Robust Approach for Brain Tumor Detection in Magnetic Resonance Images
Using Finetuned EfficientNet. IEEE Access, 10, 65426-65438.
16. Anaya-Isaza, A., &Mera-Jiménez, L. (2022). Data Augmentation and Transfer
Learning for Brain Tumor Detection in Magnetic Resonance Imaging. IEEE
Access, 10, 23217-23233.
17. Song, Y., & Ma, R. (2022). Multiple Omics Analysis of the Rac3 Roles in
Different Types of Human Cancer. IEEE Access, 10, 92633-92650.
18. Zhu, J., Teolis, S., Biassou, N., Tabb, A., Jabin, P. E., &Lavi, O. (2020). Tracking
the adaptation and compensation processes of patients’ brain arterial network to
an evolving glioblastoma. IEEE transactions on pattern analysis and machine
intelligence, 44(1), 488-501.
19. L. M. Oliveira et al., "Spectroscopic detection of pigments in tissues: correlation
with tissue aging and cancer development," 2022 International Conference Laser
Optics (ICLO), 2022, pp. 1-1,doi: 10.1109/ICLO54117.2022.9839939.
20. H. Fu et al., "HMRNet: High and Multi-Resolution Network With Bidirectional
Feature Calibration for Brain Structure Segmentation in Radiotherapy," in IEEE
Journal of Biomedical and Health Informatics, vol. 26, no. 9, pp. 4519-4529,
Sept. 2022, doi: 10.1109/JBHI.2022.3181462.
21. Kumar, E. Boopathi, and V. Thiagarasu. "Comparison and Evaluation of Edge
Detection using Fuzzy Membership Functions." International Journal on Future
Revolution in Computer Science & Communication Engineering (IJFRCSCE),
ISSN (2017): 2454-4248.
22. I. Ezhov et al., "Geometry-Aware Neural Solver for Fast Bayesian Calibration
of Brain Tumor Models," in IEEE Transactions on Medical Imaging, vol. 41, no.
5, pp. 1269-1278, May 2022, doi: 10.1109/TMI.2021.3136582.
23. X. Xu and P. Prasanna, "Brain Cancer Survival Prediction on Treatment-Naïve
MRI using Deep Anchor Attention Learning with Vision Transformer," 2022
IEEE 19th International Symposium on Biomedical Imaging (ISBI), 2022, pp.
1- 5, doi: 10.1109/ISBI52829.2022.9761515.
24. Sundari, M. S., & Nayak, R. K. (2021, November). Efficient Tracing and
Detection of Activity Deviation in Event Log Using ProM in Health Care
Industry. In 2021 Fifth International Conference on I-SMAC (IoT in Social,
Mobile, Analytics and Cloud)(I-SMAC) (pp. 1238-1245). IEEE.
25. M. A. Hafeez, C. B. Kayasandik and M. Y. Dogan, "Brain Tumor Classification
Using MRI Images and Convolutional Neural Networks," 2022 30th Signal
Processing and Communications Applications Conference (SIU), 2022, pp. 1-
4, doi: 10.1109/SIU55565.2022.9864962.
26. J. Liu, "VGG, MobileNet and AlexNet on Recognizing Skin Cancer
Symptoms," 2022 3rd International Conference on Electronic Communication
and Artificial Intelligence (IWECAI), 2022, pp. 525-528, doi:
10.1109/IWECAI55315.2022.00107.
27. M. Liu, C. Chen, Q. Yuan, L. Zhuang and P. Wang, "Olfactory Bulb Biosensor
Monitoring the Effect of Treatment Drugs for AD in vitro," 2022 IEEE
International Symposium on Olfaction and Electronic Nose (ISOEN), 2022, pp.
1-3, doi: 10.1109/ISOEN54820.2022.9789665.
28. S. M. Kurian and S. Juliet, "Savitzky–Golay Pre- processed Kulczynski
Similarity Feature Extractive Convolutional Deep Belief Network for Brain
Tumor Detection," 2022 6th International Conference on Devices, Circuits and
Systems (ICDCS), 2022, pp. 398-402, doi:
10.1109/ICDCS54290.2022.9780698.
29. I. D. Romanishkin et al., "Investigation of a combined absorption-fluorescence-
Raman spectroscopy approach to brain tumor tissue differentiation ex vivo,"
2022 International Conference Laser Optics (ICLO), 2022, pp. 1-1, doi:
10.1109/ICLO54117.2022.9840141.
30. Reddy, Ch Subba, T. L. Yookesh, and E. Boopathi Kumar. "A Study on
Convergence Analysis Of Runge-Kutta Fehlberg Method To Solve Fuzzy Delay
Differential Equations." Journal of Algebraic Statistics 13.2 (2022): 2832-2838.
31. S. Kumar Mishra, D. Kumar, G. Kumar and S. Kumar, "Multi-Classification of
Brain MRI Using EFFICIENTNET," 2022 International Conference for
Advancement in Technology (ICONAT), 2022, pp. 1-6, doi:
10.1109/ICONAT53423.2022.9725927.
32. S. S and S. Aurelia, "A Machine Learning Entrenched Brain Tumor Recognition
 Framework," 2022 International Conference on Electronics and Renewable
Systems (ICEARS), 2022, pp. 1372- 1376, doi:
10.1109/ICEARS53579.2022.9752195.
33. M. Jagadeesh, P. Chitra, K. Srilatha, M. Sumathi and I. RexilineSheeba, "Brain
Tumour Classification using CNN Algorithm," 2022 International Conference on
Electronics and Renewable Systems (ICEARS), 2022, pp. 1813- 1816, doi:
10.1109/ICEARS53579.2022.9752096.
34. K. Padmavathi, O. S. R. R. Y. Thalla, S. S. Reddy, P. Yadlapalli, T. Roshan and
T. Charan, "Transfer Learning Using EfficientNet for Brain Tumor Classification
from MRI Images," 2022 International Conference on Computer Communication
and Informatics (ICCCI), 2022, pp. 1-4, doi: 10.1109/ICCCI54379.2022.9740923.
35. L. M. Sao Khue and E. Pavlovskiy, "Improving Brain Tumor Multiclass
Classification With Semantic Features," 2022 Ural-Siberian Conference on
Computational Technologies in Cognitive Science, Genomics and Biomedicine
(CSGB), 2022, pp. 150-154, doi:
10.1109/CSGB56354.2022.9865366.
36. Kumar, E. Boopathi, and M. Sundaresan. "Edge detection using trapezoidal
membership function based on fuzzy's mamdani inference system." 2014
International Conference on Computing for Sustainable Global Development
(INDIACom). IEEE, 2014.
37. H. Li et al., "View-Disentangled Transformer for Brain Lesion Detection," 2022
IEEE 19th International Symposium on Biomedical Imaging (ISBI), 2022, pp.
1-5, doi:\10.1109/ISBI52829.2022.9761542.
38. N. Çınar, B. Kaya and M. Kaya, "Comparison of deep learning models for brain
tumor classification using MRI images," 2022 International Conference on
Decision Aid Sciences and Applications (DASA), 2022, pp. 1382-1385, doi:
10.1109/DASA54658.2022.9765250.
39. Chandra, J. V., &Pasupuleti, S. K. (2022, March). Machine Learning
Methodologies for predicting Neurological disease using Behavioral Activity
Mining in Health Care. In 2022 8th International Conference on Advanced
Computing and Communication Systems (ICACCS) (Vol. 1, pp. 1035-1039).
IEEE.
40. Y. Zhang, B. Dai, M. Dong, H. Chen and M. Zhou, "A Lung Cancer Detection
and Recognition Method Combining Convolutional Neural Network and
Morphological Features," 2022 IEEE 5th International Conference on Computer
and Communication Engineering Technology (CCET), 2022, pp. 145-149,
doi:10.1109/CCET55412.2022.9906329.
41. Yookesh, T. L., et al. "Efficiency of iterative filtering method for solving Volterra
fuzzy integral equations with a delay and material investigation." Materials
today: Proceedings 47(2021): 6101-6104.
42. M. N. de Barros, E. D. S. Souza and J. Lisboa- Filho, "Using spatial analysis to
calculate the incidence of lung cancer in the region of Muriaé- MG," 2022 17th
Iberian Conference on Information Systems and Technologies (CISTI), 2022,
pp. 1-6, doi: 10.23919/CISTI54924.2022.9820194.
43. C. Tejaswini, P. Nagabushanam, P. Rajasegaran, P. R. Johnson and S. Radha,
"CNN Architecture for Lung Cancer Detection," 2022 IEEE 11 th International
Conference on Communication Systems and Network Technologies (CSNT),
2022, pp.346-350, doi:10.1109/CSNT54456.2022.9787650.
44. Khan, S.U.R.; Raza, A.;Waqas, M.; Zia, M.A.R. Efficient and Accurate Image Classification
Via Spatial Pyramid Matching and SURF Sparse Coding. Lahore Garrison Univ. Res. J.
Comput. Sci. Inf. Technol. 2023, 7, 10–23.
45. Khan, S.U.R.; Zhao, M.; Asif, S.; Chen, X. Hybrid-NET: A fusion of DenseNet169 and
advanced machine learning classifiers for enhanced brain tumor diagnosis. Int. J. Imaging Syst.
Technol. 2024, 34, e22975.
46. Khan, S.U.R.; Zhao, M.; Asif, S.; Chen, X.; Zhu, Y. GLNET: Global–local CNN’s-based
informed model for detection of breast cancer categories from histopathological slides. J.
Supercomput. 2023, 80, 7316–7348
47. Khan, S. U. R., & Asif, S. (2024). Oral cancer detection using feature-level fusion and
novel self-attention mechanisms. Biomedical Signal Processing and Control, 95,
106437
48. Khan, S.U.R.; Asif, S.; Bilal, O.; Ali, S. Deep hybrid model for Mpox disease
diagnosis from skin lesion images. Int. J. Imaging Syst. Technol. 2024, 34, e23044.
49. Raza, A.; Meeran, M.T.; Bilhaj, U. Enhancing Breast Cancer Detection through
Thermal Imaging and Customized 2D CNN Classifiers. VFAST Trans. Softw. Eng.
2023, 11, 80–92.
50. S. A. Ali, N. Vallapureddy, S. Mannem, Y. Gudla and V. Malathy, "Detection of
Cancer in Lung CT Image Using 3D CNN," 2022 2nd International Conference
on Intelligent Technologies (CONIT), 2022, pp.1-4, doi:
10.1109/CONIT55038.2022.9848092.
51. Dai, Q., Ishfaque, M., Khan, S. U. R., Luo, Y. L., Lei, Y., Zhang, B., & Zhou, W. (2024). Image
classification for sub-surface crack identification in concrete dam based on borehole CCTV
images using deep dense hybrid model. Stochastic Environmental Research and Risk Assessment,
1-18.
52. Khan, U. S., & Khan, S. U. R. (2024). Boost diagnostic performance in retinal disease
classification utilizing deep ensemble classifiers based on OCT. Multimedia Tools and
Applications, 1-21.
53. Farooq, M. U., Khan, S. U. R., & Beg, M. O. (2019, November). Melta: A method level energy
estimation technique for android development. In 2019 International Conference on Innovative
Computing (ICIC) (pp. 1-10). IEEE
54. Shahzad, I., Khan, S. U. R., Waseem, A., Abideen, Z. U., & Liu, J. (2024). Enhancing ASD
classification through hybrid attention-based learning of facial features. Signal, Image and Video
Processing, 1-14.
55. Mahmood, F., Abbas, K., Raza, A., Khan, M. A., & Khan, P. W. (2019). Three dimensional
 agricultural land modeling using unmanned aerial system (UAS). International Journal of
Advanced Computer Science and Applications, 10(1).
56. Khan, M. A., Khan, S. U. R., Haider, S. Z. Q., Khan, S. A., & Bilal, O. (2024). Evolving
knowledge representation learning with the dynamic asymmetric embedding model. Evolving
Systems, 1-16.
57. D. Rawat, Meenakshi, L. Pawar, G. Bathla and R. Kant, "Optimized Deep
Learning Model for Lung Cancer Prediction Using ANN Algorithm," 2022 3rd
International Conference on Electronics and Sustainable Communication
Systems (ICESC), 2022,pp.889-894, doi: 10.1109/ICESC54411.2022.9885607.
58. [Wajid, M., Abid, M. K., Raza, A. A., Haroon, M., & Mudasar, A. Q. (2024). Flood
Prediction System Using IOT & Artificial Neural Network. VFAST Transactions
on Software Engineering, 12(1), 210-224.
59. Raza, A., Soomro, M. H., Shahzad, I., & Batool, S. (2024). Abstractive Text
Summarization for Urdu Language. Journal of Computing & Biomedical
Informatics, 7(02).
60. Al-Khasawneh, M. A., Raza, A., Khan, S. U. R., & Khan, Z. (2024). Stock Market
Trend Prediction Using Deep Learning Approach. Computational Economics, 1-
32.
61. [Khan, U. S., Ishfaque, M., Khan, S. U. R., Xu, F., Chen, L., & Lei, Y. (2024).
Comparative analysis of twelve transfer learning models for the prediction and
crack detection in concrete dams, based on borehole images. Frontiers of
Structural and Civil Engineering, 1-17.
62. M. Marathe and M. Bhalekar, "Detection of Lung Cancer using CT Scans with
Deep Learning Approach," 2022 7th International Conference on Communication
and Electronics Systems (ICCES), 2022, pp.1026-1031,doi:
10.1109/ICCES54183.2022.9835901.
63. K. K. Kumar and K. V. Kanimozhi, "Detection of Lung Disease using Novel
Genetic Algorithm in comparison with Particle Swarm Optimization to improve
Accuracy," 2022 International Conference on Business Analytics for
Technology and Security (ICBATS), 2022, pp. 1-4, doi:
10.1109/ICBATS54253.2022.9759044.
64. Keerthi, R.S., Dhabliya, D., Elangovan, P., Borodin, K., Parmar, J., Patel, S.K.
 Tunable high- gain and multiband microstrip antenna based on liquid/copper
split-ring resonator superstrates for C/X band communication (2021) Physica
B:Condensed Matter, 618, art. no. 413203
65. Dhabliya, D. Delay-Tolerant Sensor Network (DTN) Implementation in Cloud
Computing (2021) Journal of Physics: Conference Series, 1979 (1), art. no.
012031,