You are on page 1of 5

Early Identification and Detection of Brain

Tumor using Deep Learning Techniques

This project looks at a system that uses computational


Vivek Kumar Sharma1 methods to identify tumor blocks and classifies whether a
Department of Computer Science and Engineering
tumor is present or not. It does this by using Convolution
Parul University
Vadodara, India Dr. Gaurav Kumar Ameta2
vkshere4every1@gmail.com Department of Computer Science and Engineering
Parul University
Vadodara, India
Abstract: A brain tumor is a potentially fatal condition that gauravameta1@gmail.com
can seriously impair a patient's quality of life and interfere
with normal brain functioning. Early identification and
detection are essential for successful treatment. Deep learning
methods for image analysis and detection have been
increasingly important in the medical profession recently, Neural Networks (CNNs), a type of deep learning algorithm.
yielding more dependable and accurate findings. To detect When these algorithms are applied to MRI scan imageries,
brain cancers, segmentation is an act of differentiating between
the presence of a brain tumor can be identified with greater
normal and aberrant brain cells or tissues. In this study, we use
magnetic resonance imaging (MRI) imageries to examine accuracy and speed, which helps patients receive treatment
several methods for brain tumor detection and segmentation. more quickly.
To find anomalies, the symmetric and asymmetric shapes of the This paper aims to compare various deep learning models
brain are first analysed in the detection procedure. After that and identify the best model that can learn and get better
categorization of the cells is performed as tumors or non-
through repeated runs and parameter adjustments.
tumorigenic. The goal of this research is to develop a brain
tumor detection technique that is more precise and effective.
II. LITERATURE REVIEW
For detection and classification, different deep learning models
like DenseNet201, CNN, U-Net, and U-Net++ are applied to the A deep learning model for automated segmentation and
MRI image dataset. As the result for classification DenseNet classification of brain tumors is proposed [3]. Based on
shows best accuracy 98.11% and for segmentation of tumor convolutional neural network (CNN) architecture, the model
UNet++ shows 98% accuracy.
can identify whether a brain Tumor is benign or malignant
Keywords- Brain-Tumor detection, Deep Learning, by dividing it into distinct regions. Using an MRI scan
Convolutional Neural Network (CNN), DenseNet, Magnetic dataset as a test dataset, the model produced segmentation
Resonance Imageries and classification accuracy of 96.2% and 97.4%,
I. INTRODUCTION respectively.
A multimodal deep learning method was suggested for
The health and wellness sector differs greatly from other classifying and segmenting brain tumors [4]. To increase
sectors of the economy. People always demand a high segmentation and classification accuracy, the method
degree of awareness and comfort, when it comes to the brain combines a long short-term memory (LSTM) network with a
[1], it is an essential and highly sensitive organ that powers convolutional neural network (CNN). Using an MRI scan
the body, other organs, and facilitates decision-making. It dataset as a test dataset, the model produced segmentation
functions as the central nervous system's (CNS) control and classification accuracy of 98.6% and 99.1%,
centre. The brain is the controller of the body's daily respectively.
voluntary and involuntary functions. A brain tumor detection method was developed by using
In most cases, a brain tumor can result in serious health both intelligent and automatic system using Lee sigma
conditions like cancer because it resembles a mesh of filtered histogram segmentation model [5]. Lee sigma
harmful and undesired tissue growth in the brain. Brain filtered histogram algorithm used for detection of tumor in
failure results from brain tumor cells growing to the point brain in minimum time. The tanh activation function used
where they eventually absorb all the nutrients meant for the for classification purpose. Gray bimodal histogram
healthy cells and tissues. For the patient, the treatment, and segmentation process for measuring Jaccard index for
even the patient's survival, the tumour’s identification is similarity among pixels.
essential [2]. To manually identification for Tumors, medical An algorithm was proposed for the detection of brain tumor
professionals use computed tomography (CT) and magnetic from MRI imageries using deep convolutional neural
resonance imaging (MRI) techniques. Being able to treat a network [6]. Tumor is detected from two different classes
brain tumor in its early stages before it becomes malignant with different models i.e., binary class and multiclass.
depends on early detection of its existence. Brain tumors are classified using a capsule network because
as per author it is a better option as compared to CNN. CNN
might lose track of information while processing it. To

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE


increase classification accuracy, the Bayesian approach was convolutional network architecture. It is used in various
applied to the capsule network [7]. The benchmark brain medical imaging applications [8]. This is a U shape
tumor dataset was used, and the BayesCap framework was architecture, which consists of encoding and its
required for model testing. corresponding decoding path. There are three steps in the U-
Net: (1) encoder (2) bottleneck (3) decoder. At the encoder
III. METHODOLOGY step, the input image is passed through a series of max
In this paper, segmentation methods like U-Net, U-Net++ pooling and the ReLU activation algorithm. It comes after
are used for segmentation of the tumor from MRI imageries, the convolutional layer to reduce the dimension of image
whereas Convolutional Neural Network (CNN) and and to increase the features from the input image.
DenseNet are applied for the classification of tumor. These
methods are trained on MRI imageries dataset. Performance The bottleneck step emerges at the bottom of U-Net,
was measured using performance matrix through the containing higher number of features and information about
parameters like accuracy, loss value, and recall value. The the input image. Whereas, the decoder step concatenates the
performance matrix result assists in finding of best method features from the encoder block using skip connections for
by comparing accuracies of all the methods.
increasing number of features at every stage for accurate
A. Dataset image segmentation.
A dataset containing 3064 Magnetic Resonance Imaging Figure 2 shows the architecture of U-Net. In this architecture
(MRI) of brain imageries was considered for investigation. input image passes through two-dimensional convolutional
Imageries of meningiomas, gliomas, pituitary gland tumors, layer with 32 convolution filters, 3×3 kernel size, followed
and healthy brains were used for brain tumor detection from by size of 2×2 max pooling layer. After that input image is
MRI imageries. After collecting all imageries, masked passed into the process of encoder and decoder for finding
imageries were created manually using label-studio tool and output. The output layer uses the kernel of size 1×1, number
all imageries were saved as .png format. The masked of classes and sigmoid function as activation function. After
imageries are now added into dataset of Brain Tumor MRI Figure 2. U-Net Model
imageries. This dataset can be retrieved from github
repository and it is publicly available on ( https://github.com/MRI- defining the model, training of model is done for finding out
brain-tumor-detection/tree/main/image_dataset). the training parameters, loss, accuracy, and validation
. results.
B. Data Pre-Preprocessing
2) U-Net++
All the imageries of the dataset are of different formats and
sizes. It will display an error when it reads different images
as an input at the time of training. To solve this problem, all U-Net++ architecture is an extension of U-Net architecture
the imageries in the dataset are converted into same size of as shown in Figure 3. This nested architecture is created
256×256 pixels For pre-processing, in the first step all the recursively by connecting the skip connection of U-Net.
imageries are converted into same image format, here all This process allows the model to extract more features from
mask imageries are in .png format and original imageries are the images of different resolution and scale. It helps in
in different image formats. In the second step, all the accurately segmenting the object from the image [9]. This
imageries (original as well as mask imageries) are converted has same encoder and decoder block as in U-Net, but there
into gray scale image format for saving time. Now, all is a nested connection between the encoder and decoder
imageries are of 256×256×3 converted into 256×256×1. block at different resolutions. This nested feature in encoder
Figure 1 shows the original image of size 256×256×3 of and decoder block gives more accurate features and
healthy and tumorous imageries, gray scale imageries of the accuracy.
sizes 256×256×1 and mask imageries of heathy and
tumorous.

C. Deep Learning based segmentation techniques for Brain


tumor detection
There are number of algorithms available for brain tumor
segmentation on MRI imageries like U-Net, U-Net++. In
this paper, segmentation techniques are used for extracting
tumor from MRI imageries. In all above algorithms
mentioned, different layers consist of convolution block,
encoder, and decoder block. These encoder and decoder
blocks consist of number of filters, pooling layer and skip
connection for segmenting tumor accurately. After
segmentation, performance matrix is calculated for finding
best algorithm for brain tumor segmentation. Figure 3. U-Net++ model

1) U-Net algorithm
3) Convolutional Neural Network
The U-Net algorithm is a deep learning segmentation
algorithm and its architecture is fully connected
A type of artificial neural network (ANN) frequently Comparison of the different methods and algorithms used
employed in picture analysis and recognition. The visual for segmentation and classification purposes are discussed in
cortex, the area of the brain responsible for processing of this section. The number of deep learning algorithms like U-
visual data, serves as an inspiration for the design and Net, U-Net++, DenseNet and CNN are discussed in this
operation of CNN. paper. This section compares the performance of model
A sequence of filters is applied to an input image by CNNs explored in this paper. On training of these entire models on
to function. A certain area of the image is subjected to a the brain imageries dataset, different accuracy loss and
filter, which is essentially a tiny weight matrix. Each filter values were achieved. Table 1 and Table 2 represents
produces a feature map, or a map of the filter's activation accuracy and performance analysis respectively. The table 1
values throughout the image. After going through several shows the accuracies of different models like for
pooling layers, the feature maps lose some of their size and classification CNN gives 96%, DenseNet gives 98%,
become more resilient to noise and changes in the input accuracy while for segmentation of tumor UNet and UNet+
image [10]. + gives 76% and 98% respectively.
A fully connected layer receives the output from the pooling
layers and uses it to categorize the image. Backpropagation TABLE1. Comparison of accuracy analysis
is a technique used to train the fully connected layer. It
modifies the filter and pooling layer weights into reduces the Validation Validation
discrepancy between the image's actual and expected Model Accuracy Loss
Accuracy Loss
classification.
CNN 0.9733 0.09 0.9600 0.051
4) DenseNet DenseNet 0.9700 0.023 0.9811 0.032
U-Net 0.8621 0.010 0.7633 0.033
The CNN architecture known as DenseNet, or Dense
Convolutional Network, has demonstrated encouraging U-Net++ 0.9583 NAN 0.9760 NAN
outcomes in the identification of brain Tumors [11]. The
capacity of DenseNets to extract intricate patterns and TABLE2. Performance analysis of models
features from medical imageries makes them especially
suitable for this kind of work. The architecture of DenseNet Method Precision Recall F1 score
is shown in Figure 4.
CNN 0.49 0.48 0.49
DenseNet 0.49 0.47 0.48
U-Net 0.53 0.67 0.33
U-Net++ 0.88 0.93 0.90

Figure 5 shows Graphical representation of training


accuracy versus epochs of CNN model, Figure 6
Graphical representation of Loss accuracy versus epochs
of CNN model, Figure 7 Graphical representation of
training accuracy versus epochs of DenseNet model,
Figure 8. Graphical representation of Loss accuracy
versus epochs of DenseNet model, Figure 9 Graphical
representation of training accuracy versus epochs of
UNet model, Figure 10. Graphical representation of Loss
versus epochs of UNet model and Figure 11. Graphical
representation of training accuracy versus epochs of
UNet++ model and Loss versus epochs of UNet++
model.

Figure 4. DenseNet Architecture

Based on two situations, the brain tumor was graded by


cropping the tumor lesions, and uncropped brain imageries.
CNN architecture is used rather than using transfer learning
techniques and pre-trained CNN like DenseNet201.
IV. RESULT ANALYSIS
Fig 8. Graphical representation of Loss accuracy versus epochs of
DenseNet model
Fig 5 Graphical representation of training accuracy versus epochs of CNN
model

Fig 9 Graphical representation of training accuracy versus epochs of


UNet model
Fig 6 Graphical representation of Loss accuracy versus epochs of CNN
model

Fig 10. Graphical representation of Loss versus epochs of UNet model

Fig 7 Graphical representation of training accuracy versus epochs of


DenseNet model
[3] AL-HILO, R. U. A. A., and MASHKOOR MAHMOOD, "BRAIN
TUMOR SEGMENTATION AND CLASSIFICATION USING
MACHINE LEARNING." PhD diss., 2023.

[4] Li, Yuexiang, and Linlin Shen, "Deep learning based multimodal
brain tumor diagnosis." In Brainlesion: Glioma, Multiple Sclerosis,
Stroke and Traumatic Brain Injuries: Third International Workshop,
BrainLes 2017, Held in Conjunction with MICCAI 2017, Quebec
City, QC, Canada, September 14, 2017, Revised Selected Papers 3,
pp. 149-158. Springer International Publishing, 2018.

[5] Kurian, Simy Mary, and Sujitha Juliet, "An automatic and intelligent
brain tumor detection using Lee sigma filtered histogram
segmentation model." Soft Computing 27, no. 18 (2023): 13305-
13319.

[6] Khan, Md Saikat Islam, Anichur Rahman, Tanoy Debnath, Md Razaul


Karim, Mostofa Kamal Nasir, Shahab S. Band, Amir Mosavi, and
Iman Dehzangi, "Accurate brain tumor detection using deep
convolutional neural network." Computational and Structural
Biotechnology Journal 20 (2022): 4733-4745.
Fig 11. Graphical representation of training accuracy versus epochs of
UNet++ model and Loss versus epochs of UNet++ model. [7] Afshar, Parnian, Arash Mohammadi, and Konstantinos N. Plataniotis,
"Brain tumor type classification via capsule networks." In 2018 25th
IEEE international conference on image processing (ICIP), pp. 3129-
V. CONCLUSION 3133. IEEE, 2018.
It is essential to detect brain Tumors accurately, since
[8] Ronneberger, Olaf, Philipp Fischer, and Thomas Brox, "U-net:
misidentification can result in an irreversible illness. The Convolutional networks for biomedical image segmentation."
medical team in neurology department mostly remain In Medical Image Computing and Computer-Assisted Intervention–
overburdened due to a considerable increase in patients and MICCAI 2015: 18th International Conference, Munich, Germany,
an increasing likelihood of brain tumors. Consequently, to October 5-9, 2015, Proceedings, Part III 18, pp. 234-241. Springer
International Publishing, 2015.
solve these problems, an accurate and effective brain tumor
detection approach is desperately needed. The segmentation
[9] Zhou, Zongwei, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh,
and classification algorithms are used to segment and and Jianming Liang, "Unet++: A nested u-net architecture for medical
classify tumor from the MRI brain imageries by dividing the image segmentation." In Deep Learning in Medical Image Analysis
whole image into small pixels and extract all the relevant and Multimodal Learning for Clinical Decision Support: 4th
International Workshop, DLMIA 2018, and 8th International
features from the brain imageries. These algorithms help Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018,
professionals in finding out the tumor in particular region in Granada, Spain, September 20, 2018, Proceedings 4, pp. 3-11.
lesser time with more accuracy. Segmentation plays an Springer International Publishing, 2018.
important role in classification of tumor. After performing
segmentation, the accuracy of classification becomes higher. [10] Pavani, D., Durgalaxmi, K., Datta, B. S., & Nagajyothi, D., “Brain
By observing results of these algorithms on the same Tumor Detection Using Convolutional Neural Networks in MRI
Imageries”. In International Conference on Soft Computing and
dataset, the U-Net++ deep learning model gives higher Signal Processing (pp. 751-760). Singapore: Springer Nature
accuracy as compared to the U-Net for segmenting tumor Singapore, (2021, June).
from brain imageries.
[11] Özkaraca, Osman, Okan İhsan Bağrıaçık, Hüseyin Gürüler, Faheem
REFERENCES Khan, Jamil Hussain, Jawad Khan, and Umm E. Laila.,"Multiple
Brain Tumor Classification with Dense CNN Architecture Using
Brain MRI Images." Life 13, no. 2 (2023): 349.
[1] Sadad, Tariq, Amjad Rehman, Asim Munir, Tanzila Saba, Usman
Tariq, Noor Ayesha, and Rashid Abbasi, "Brain tumor detection and
multi‐classification using advanced deep learning techniques." [12] Mahmud, Md Ishtyaq, Muntasir Mamun, and Ahmed Abdelgawad, "A
Microscopy Research and Technique 84, no. 6 (2021): 1296-1308. deep analysis of brain tumor detection from mr images using deep
learning networks." Algorithms 16, no. 4 (2023): 176.
[2] Dipu, Nadim Mahmud, Sifatul Alam Shohan, and K. M. A. Salam,
"Deep learning based brain tumor detection and classification." In [13] Asad, Rimsha, Saif Ur Rehman, Azhar Imran, Jianqiang Li, Abdullah
2021 International conference on intelligent technologies (CONIT), Almuhaimeed, and Abdulkareem Alzahrani, "Computer-Aided Early
pp. 1-6. IEEE, 2021. Melanoma Brain-Tumor Detection Using Deep-Learning Approach."
Biomedicines 11, no. 1 (2023): 184.

You might also like