Brain Tumor Detection and Segmentation using UNET

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 09 | Sep 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 589
Brain Tumor Detection and Segmentation using UNET
Kunal S. Tibe1, Abhishek Kangude2, Pratibha Reddy3, Pratik Kharate4, Prof. Vina Lomte5
1,2,3,4Final Year, Dept. of Computer Engineering, R.M.D Sinhgad School of Engineering, Maharashtra, India
5H.O.D, Dept. of Computer Engineering, R.M.D Sinhgad School of Engineering, Maharashtra, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Brain tumor detection is one of the most complex
biomedical problems. Its anatomical structure makes it
complex to cure neuro medical issues. Medicalsegmentationis
a challenging part in curing intense brain tumors. In such
scenarios, deep learning algorithms are used to resolve the
complexity of segmentation and detect the tumor accurately.
The ConvolutionalNeuralNetworks(CNN)hasbeen developed
by the efficient auto segmentation technology. UNET is used
along with methods of computer vision to increase the rate of
successfully detecting the tumor. Use of biomedical image
segmentation extensively has resulted in high rates of curing
the tumor accurately. In this paper, we are proposing a
multimodal brain tumor segmentation using 3D UNET. We
have used the BraTS 2020 dataset which contains 369 3DMRI
images that are used for training while 125 MRI images that
are used for testing. We have developed a 3D model which
generates the output in 3D format and were able to achieve
accuracy of 98.5%.
Key Words: Brain Tumor, BraTS 2020, CNN, Gliomas, MRI,
UNET, Segmentation.
1. INTRODUCTION
Brain cancer treatmentrequireshighexpertiseandprecision
as it is one of the most complex proceduresinthe biomedical
field. Deep learning is a class of machine learning. Deep
learning algorithms use multiple layers in order to solve
complex problems. In this paper, we have used methodslike
CNN: UNET to performsegmentationoverthetumoraffected
area of the brain. The main problem of detecting a brain
tumor is its area of infection as well as the intensity. Hence,
brain tumor segmentation is done. For image segmentation,
the method of image inpainting is used. Imageinpaintingisa
method done prior to UNET in order to enhance the image
quality. It helps to smoothen the image. This helps in
identifying the tumor more accurately with the help of
similarity index. Using the 3D MRI images given in the
dataset, we have imported them to the system to perform
UNET and perform segmentation to acquire the accurate
tumor present in the brain. Mathematical models such as
Edge-Based Detection Method, Rough Set-Based Fuzzy
Clustering and Cross-EntropyLossFunction.Withthehelpof
these models, segmentation is done precisely.
2. LITERATURE REVIEW
Various researches clearly depict the importance of deep
learning methods for learning and identifying specific
patterns for accurate characterization of a specific problem.
In the Recent study performed byLingTanandWenjieMa[1]
within the context of medicalimage segmentationintheyear
2021 they have proposed a multimodal brain tumor image
segmentation method based on ACU-Net network using
different volumes of datasetsuchasBraTS2015,BraTS2018,
BraTS 2019. This study definitely showed an increased dice,
recall, precision and can employ an active contour model to
overcome image noise and edge slits to better identify low-
contrast and low-resolution in the image but This study is
also presented with poor adaptability.
As we know that gliomas are malignant and heterogeneous
Mahnoor Ali,Asim Waris proposed an ensemble of two
segmentation networks 3D CNN and a U-net [2] which helps
us for better and accurate predictions. In this study
researchers used BraTS 2019challengedatasetandachieved
a good dice score of 0.91 for whole tumor region and
provided efficient and robust tumor segmentation across
multiple regions. There were certain discrepancies and
deficiency as the recognition rate was only 83.4% hence
better pre and post processing of data was required.
The most formal and important step for brain tumor
segmentation is to detect the boundary of the tumori.e.,edge
detection. A group of researchers named Ahmed H. Abdel-
Gawad, Lobna A. Said proposed a System for Optimized edge
detection in case of brain tumor detection by analyzing MRI
images [3]. It uses different Techniques and Algorithms such
as Balance contrast Enhancement Technique (BCET), skull
stripping and Genetic Algorithm (GA) which comes under
evolutionary sciences.
For proper diagnosisandplanningforbraincancertreatment
proper brain tumor segmentation is very important so in the
year 2020 certain researchers named Nagwa M. Aboelenein,
Piao Songhao, Alam Noor proposed an architecture called
Hybrid Two Track U-net (HTTU) which is usedforAutomatic
Brain TumorSegmentation[4].Itusesdifferentmethodssuch
as N4ITK Bias Correction, Focal loss and Generalized dice
score functions. It uses BraTS2018datasetandalsoachieved
0.865 dice score for whole tumor but it could not identify the
underlying layers of the images used.
By using featurefusion on several levels wecanmakefulluse
of hierarchical features to reveal the importance of
refinement and aggregation of features in brain tumor
segmentation.Dongyuan Wu,Yi Ding,Mingfeng Zeng [5]
developed a Multi Features Refinement and Aggregation

(MRA) Model for improvement of segmentation accuracy. It
uses BraTS 2015 dataset and achieves 0.83 dice score for
whole tumor but has poor performance in case of enhancing
tumor.
Yibo Han and Zheng Zhang developed an Image Interactive
framework for Brain Image Segmentation [6] which is
assisted by Deep Learning concepts. It uses Deep
Convolutional Neural Networks (DCNN) which is a Multi-
Task Deep Learning Approach (MTDLA) but it produces
minor mistakes due to unsupervised fine tuning.
During Brain Tumor Segmentation based on MRI Modalities
certain noisy regions affect the accuracy and performance
greatly. Guohua Cheng and Hongli Ji developed an system
which is used to evaluate Adversarial Perturbation based on
MRI Modalities [7] but when four modalities are attacked or
damaged we will observe a severe degradation of
performance and accuracy will occur.
To improve the performanceof medical image segmentation
and enhancement of local feature extraction Jianxin Zhang,
Zongkang developed an Attention Gate ResU-net model for
Automatic Brain Tumor Segmentation [8]. They used
different dataset volumes such as BraTS 2017,2018 and
2019. It surely outperform baselines of U-net and ResU-net
but loses an amount of context and local details among
different slices.
For proper extraction of Image features in brain tumor
segmentation Weiguang Wang,Fanlong Bu developed an
explanatorymodelwhichcombinesLearningMethodsofCNN
with feature extraction of Images in Brain Tumor detection
[9]. It is used for performance analysis, diagnostic reports
and as theoretical references for any related research.
Dataset used is GBM data volumes and it achieves 0.872 dice
score for whole tumor but has high computational time and
complexity.
For better Supervision and Excitation function Ping Liu, Qi
Dou, Qiong Wang developed an Encoder-Decoder Neural
Network [10] containing 3D Squeeze and Excitation andalso
including Deep Supervision for Brain Tumor Segmentation.
They used BraTS 2017 dataset and N4BiasFieldCorrection
algorithm along with V-net (DSSE-V-net). Certaindrawbacks
were found such as limited size kernel and receptive field
problem.
There are various types of neural networks which can be
used for Brain Tumor Segmentation, Wu Deng, Qinke Shi
proposed a Deep Learning based model for brain tumor
segmentation which uses two neural networks
Heterogeneous ConvolutionalNeuralNetworks(H-CNN)and
CRF-RecurrentRegressionbasedNeuralNetwork(RNN)[11].
It uses the BraTS 2017 dataset for evaluation and achieves
high accuracy and recognition rate.
Tamjid Imtiaz, Shahriar Rifatdeveloped an Automated Brain
Tumor Segmentation system which is based on Multi-Planar
Superpixel Level Feature Extraction [12]. To reduce the bias
in intensities an Intensity Adjustment Scheme is applied on
the whole 3D MRI images. Extremely Randomized Trees are
used for classification of tumor and non-tumor classes. This
study uses NCI-MICCAI 2013 challenge dataset with certain
drawbacks found such as low level of precision in the tumor
region segmentation.
As most medical imaging dataset which are used for brain
tumor segmentation are small and fragmented Changhee
Han, LeonardoRundo,RhyosukeArakidevelopedasystemas
Brain Tumor Augmentation for Tumor detection which is
achieved by combining Noise-to-Image and Image-to-Image
GAN [13]. It uses BraTS 2016 dataset with ResNet-50 and t-
SNE methods but results in poor optimization results.
The classification of brain cancer by manualmethodrequires
the knowledge and expertise of a Physician who is
experienced enough in the medical field. Abdu Gumaei,
Mohammad Mehedi Hassan came up with a Hybrid Feature
Extraction Method for Brain Tumor Segmentation that
includes implementation of Regularized Extreme Learning
Machine (RELM) [14]. BraTS 2013 dataset, Fuzzy C-means
algorithm and Cellular Automata with a Grey Level Co-
occurrence matrix (GLCM) are used in this study and
relatively low efficiency is present.
Cascading is a very popular concept in neural networks
classes. Kai Hu, Qinghai Gan,Yuan Zhang developed asystem
for brain tumor segmentation which uses Multi- Cascaded
Convolutional Neural Networks (MC CNN) and Conditional
Random Field (CRF) for its implementation [15]. When
integrating into 3D CNN Poor effectiveness of images is
observed.
Out of the multiple imaging techniques used to detect brain
tumors MRI is commonly used as it has superior image
quality with highest resolutions are obtained without any
radiation. In the year 2019, Hossam H. Sultan, Nancy M.
Salem, Walid AL-Atabany proposed a Deep Learning model
which is based on multi-classification method for Brain
Tumor Images [16]. It uses Nanfang Hospital and General
Hospital and The Cancer Imaging Archive (TCIA) public
access repository for evaluation of the proposed model.
Aimin Yang, Xiolei Yang, Wenrui Wu did a research study in
the year 2019 based on Feature Extraction of Tumor Image
using Convolutional Neural Networks (CNN) [17] which
concluded that CNNalgorithm shows highaccuracy intumor
image feature extraction and also demonstrated different
advantages of CNN in neuroimaging field. In this study local
binary model algorithm and convolutional neural network
algorithm are used to extract the required features of tumor.

In recent years the applications of AI in Magnetic Resonance
Imaging (MRI) have been applied in various biomedical
studies. In the year 2018 Gunasekaran Manogaran, P.
Mohamed Shakeel developed a Machine Learning Based
Approach for Brain Tumor Detection and data sample
imbalance analysis [18]. It contains an improved orthogonal
gamma distribution-based method which is used to analyze
the under-segments and over-segments of brain tumor
regions to automaticallydetectabnormalitiesintheRegionof
Interest (ROI) but it has low real-time applications.
In the neuroscience field, different Algorithms are used for
precise andaccuratebrain tumor Segmentation.QingnengLi,
Zhifan Gao proposed a Unified Algorithm for Glioma
Segmentation using Multimodal MRI images [19]. It uses a
Two Step Refinement Strategy to maintain PPV values which
in turnincreasesprocessingtimeandreducestherecognition
rate.
Guotai Wang, Wenqi Li, Maria A. Zuluaga developed an
Interactive Medical Image Segmentation modelwiththehelp
of Image-Specific Fine Tuning [20]. This model uses Pre-
trained Gaussian Mixture Model (GMM) so it has fewer user
interactions and less user time than traditional interactive
segmentation methods. But this model requires a large
number of annotated images for training.
1.1 ABBREVIATIONS AND ACRONYMS
CNN (Convolutional Neural Network), BraTS (Brain Tumor
Segmentation), MRI (Magnetic Resonance Imaging), CRF
(Conditional Random Fields), GAN (Generative Adversarial
Networks), PPV (Positive Predictive Value).
3. LIVE SURVEY
The survey discussed here aims for the live impact of other
methodologies in Brain Tumor segmentation earlier. In this
survey, hospitals which are currently working on detection
and segmentation of Brain Tumor based on this particular
method gives us a clear idea about the procedureand overall
desired results acquired during their process.
In order to identifythesub-regionssuchasEdema,enhancing
tumor, and Necrosis this Tata Memorial Hospital proposed a
3D UNET architecture which helped in identifying these
radiological sub-regions of a Brain Tumor. To keep the
balance between the tumor and non-tumor patches inside a
brain, patch extraction scheme has also been put forward to
address the problem. And along with which the architecture
helps for precise location of a Tumor by using symmetric
expanding featuresand to capture the contextbycontracting
path. The desired results achieved during the process
contains the Dice Scores of 0.92, 0.90 and 0.81 for Whole
Tumor (WT), Tumor Core (TC) and Enhancing Tumor (ET),
respectively based on independentpatients’datasetfromthe
hospital.
Present work in Apollo Hospital gives insight about a visual
and quantitative analysis on automated Brain Tumor
segmentation using Fuzzy-Possibilistic C-Means (FPCM)
methodology. The K-means w.r.t patch-based technique is
also used to carry out the skull scripting,alongwithRegionof
Interest (ROI) it is possible to measure the quantity of the
exact region of a brain tumor located. The results involved
are improved based on performance achieved than the
previous algorithms used to carry out the same benchmark
results.
4. DATASET DETAILS
4.1 PREPARATION
The dataset used for our system is BraTS2020 Dataset
(Training+Validation).ItcontainsMulti-MagneticResonance
Imaging (MRI) scans. The Dataset mainly focuses on the
segmentation of an essential heterogeneous brain tumor,
namely gliomas. It is also used to evaluate the uncertainty
between different algorithms in tumor segmentation.
 Dataset Reference - www.kaggle.com
 Size - 40GB
 No. of Training Sets - 369
 No. of Validation Sets – 125
4.2 PREPROCESSING
Before loading the dataset into thetrainingmodel,wehaveto
perform data pre-processing operations to remove noisy
regions and extract key characteristics required for the
segmentation. BraTS 2020 does not contain any missing
values or set of MRI images for any patient and all the data
fields are well labelled. The dataset is split into two parts:
Training and Testing for further evaluation.
5. PROPOSED WORK
With the survey done above there are certain remarkable
discrepancies and inefficiencies with regards to the
segmentation accuracy which found to be low due to lack of
efficient algorithms. Pre-processing of the data is not done
due to which the dataset might contain noisy regions and
erroneous values in-turn affecting the accuracy and
recognition rate. Along with which boundary and depth
detection of tumors is notaccuratewhich makesitdifficultto
perform necessary actions. Also, the process is semi-
automated with a low performance index that involves
manual work. Thus, poor results obtainedwithregardstothe
enhancing tumor region. Our proposed system overcomes
certain problems identified in the previous work. With this
proposed system, we aim to increase the accuracy of the
Brain Tumor segmentation by using a better training set and
algorithmic patterns forthe Whole Core(WC)andEnhancing

tumor region. Projecting the Tumor region is segmented in
3D format for better understanding of underlying layers,
depth, location, size and length. Also, better preprocessing
and post-processing of data has been done for efficient
output.
5.1 SYSTEM ARCHITECTURE
After the completion of Requirement gathering and
Requirement Analysis, we arrive at the design stage of our
implementation. The System Architecture represents the
overall functionalities, modules, process flow, and a brief
overview of the complete system through the eyes of a
developer. Different modules have different assigned
functionalities which are implemented in a defined manner
with supervised control flow. In our proposed system we
have converted the tumor detection operation into small
modules which increases modularity and also strikes out
various interdependencies present in case of simultaneous
execution. Fig. 1 shows the System architecture of our
System.
Fig. 1 – System Architecture
The dataset used contains a set of 3D MRI images of every
individual patient which are used as input for our system.
The main functional block is named as “Training + Testing”
which contains different operational modules which are
interlinked and follows a sequential pattern for its
implementation. First sub-module is used for Data
Preprocessing which includes operations such as Data
cleaning, Data Transformation etc. Thenextmodulefollowed
by Data preprocessing is Feature Extraction, it contains
different operations such as skull stripping, boundary
detection algorithms, cluster analysis, intensity, slicing
methods etc. which are used for accurate delineation of
Gliomas.After the completion of the first two sub- processes,
we provide the outcome as input images to the U- net
Architecture which contains a defined set of convolutional
layer information and algorithmic conditions specified for
segmentation. The outcome received after execution of the
“Training + Testing” functional block we provide this as an
input for Data Post-processing which provides us only the
segmented portion of brain tumor as an output.
5.2 UNET ARCHITECTURE
U-Net isan advanced versionofasophisticatedconvolutional
neural network (CNN) that was specially developed for
biomedical image segmentation. The feature about U-Net
which differs from CNN is that thetargetisnotonlytoclassify
whether there is an infection or not but also to identify the
area of infection. U-Net yields more precise segmentation by
using comparatively fewer high-resolution images than
previously used FCN (Fully Convolutional Networks) and
other segmentation models. It also does not require multiple
runs and a labelled set of images for segmentation. U-Net is
divided into three main layers namely, Contraction Path,
Bottleneck and Expansion Path. In order to work on fewer
parameters Max Pooling operation is carried out which
reduces the size of the feature map. In this process, the
popular activation functionReLU(RectifiedLinearActivation
Function) has been used which outputs the positive input
directly but outputs zero if the input is negative.
Fig. 2 – UNET Architecture
5.3 MATH
For implementation of any algorithmic model certain
mathematical constraints must be satisfied to ensure proper
structure and provide strong calculative measures for
productive assessment and evaluation of outcomesreceived.
5.3.1 EDGE BASED DETECTION METHOD
Edge based detection itself includes multiple mathematical
methods that identify the coordinate pixels of an MRI scan at
which the imagebrightnessmightshowsomediscontinuities.
The pixels at which the brightness changes sharply are
clustered into a set of curved line segments called edges. In
case of Brain Tumor Segmentation, we can drastically
improve the performanceofanyneuralnetworkappliedifwe
first detect the boundaryofthewholetumorregioninsteadof
applying the network on a complete MRI image. If we apply
U-net architecture on a complete brain MRI image then the
processing time will be considerably high, the recognition

rate will be very low due to the saturation in the
convolutional layers after repetitive scans.
𝑎𝑟𝑔𝑚𝑖𝑛𝑢 𝛾∥❑𝑢∥0 + ∫(𝑢 + 𝑓)2 𝑑𝑥
…Eq. 1
…Eq. 2
…Eq. 3
5.3.2 ROUGH SET-BASED FUZZY CLUSTERING
A rough set-based fuzzy clustering consists of two steps,
initialclustering based onroughsetandsecondaryclustering
based on fuzzy equivalence relations. The RSFCL algorithm
has preferable clustering validity and high run efficiency in
handling the clustering problems of both numerical data and
nominal data. In case of Brain Tumor Segmentation after
successfully detecting the tumor boundary by using edge-
based detection method, we use rough set to locate pixels
which are included in the isolated region formed after Edge
detection and Fuzzy Clustering to form clusters of pixels on
the basis of similarity in intensityvariation.Thisclusteredset
of pixels also resides in the region confined inside the tumor
boundary.
…Eq. 4
5.3.3 CROSS-ENTROPY LOSS FUNCTION
Cross entropy loss is also called logarithmicloss or logloss,it
is used along with U-net Architecture to minimize the loss
occurred during training. Cross entropy is the most
commonly used loss function in machine learning as it
measures the performance of a classification model by
considering output as a probability value between 0 and 1.
𝐻 (𝑝, 𝑞) = − 𝐸𝑝 [𝑙𝑜𝑔 𝑞]
…Eq. 5
…Eq. 6
6. RESULTS AND DISCUSSION
6.1 RESULTS
In our proposed system, we have successfully extracted the
segmented brain tumor images. We adopted the method of
UNET to extract the segmented tumor in the format of two-
dimensional(2D)aswellasthree-dimensional(3D).Thebase
architecture used is the same for both methods 2D & 3D. In
this section working functionalities, evaluation metrics,
performance scores, graphical representation of trends in
case of parameters used are shown in a detailed manner. To
measure the segmentation effect of UNET, this paper adopts
subjective visual evaluation as wellas manualcalculationsto
judge the experimental results. The results section will be
divided in two parts for precise articulation of outputs
achieved for our system.
6.1.1 2D EVALUATION
6.1.1.1 MATERIALS AND METHODS
In our system we have used Python 3.7 as the programming
language throughout the modules as it provides a vast
support of libraries and also increases productivity. For
training the deep learning model we have used keras as a
backend with TensorFlow support. Keras contain various
layers which are basic building blocks of neural networks.
From layers certain functionalities are imported such as
BatchNormalization,ActivationfunctionandConvolutionetc.
Adam optimizer is used for the training model as it is the
most efficient optimizerincaseofconvolutionaryoperations.
For fitting the training model, the batch size is kept as 32 and
epoch are set to 40. After the completion of training, we save
the weights of the trained model in HDF5 (.h5) format for
further evaluation.
6.1.1.2 OUTPUTS AND PERFORMANCE SCORES
With the proposed architecture, the segmented tumor is
extracted in the form of 2D images. Fig. 3 shows the image of
tumor extracted.
Fig. 3 – 2D Output
 RecognitionRate(Sensitivity/Recall):Itisdefined
as a measure on how well a system can identify true

positive values. It is also named as TRUE POSITIVE
VALUE. Our system has achieved a recognition rate
of 96.13 %.
…Eq. 7
 Specificity: Specificity is the measure of finding the
proportion of actual negative values which are
predicted negative. Hence, it is also called TRUE
NEGATIVE VALUE. The specificity produced for the
proposed system is 99.91 %.
…Eq. 8
 Precision: Precision is the ratio between the True
positive and all the positives. Amount of precision
recorded is 92.30 %.
…Eq. 9
 Accuracy: Accuracy is defined as the number of
correctly predicted values from all the data values.
The overall accuracy for the proposed system is
measured as 99.10 %.
Fig. 4 shows the constant and minimal increase in accuracy
over every subsequent epoch cycle.
Fig. 4 – Accuracy vs. Epoch
6.1.2 3D EVALUATION
6.1.2.1 MATERIALS AND METHODS
For implementation of the model proposed which generates
the output result of segmented tumor in 3D format there are
certain prerequisites that we have to fulfill before jumping
into the coding stage. These prerequisitescan be categorized
as software requirements and hardware requirements
depending upon the platform used for implementation. In
case of Deep learning models, a strong processor capable of
complex computation is required to reduce the processing
time and increase the efficiency along with accuracy. In the
general scenario, GPUs with CUDA enabled processors are
used for training of neural networks and machine learning
models. Certain computations can be performedonCPUwith
moderate RAM before training the model. The
hyperparameters can be manipulated for various runs to
capturedifferent behavior of the model as trainingthemodel
to achieve high accuracy and precision may require trial and
run again approach before, we overfit or underfit the model.
When it comes to software requirements and specification
the compatibility and supportability of programming
language along with IDE used is very important. Also, if
someone desires to train a particular neural model on GPU
forbetter performance then CUDAcompatibilityandsupport
for present NVIDIA drivers must be checked and verified
before gettingintoanyfurtherprocesses.Nowadaysthereare
various options to integrate our system on a cloud-based
architecturewhich diminishes highcomputationalhardware
requirements completely.
In 3D evaluation, we have used Python 3.6.3 language for
programming for individual components. Also, for neural
networks we have used PyTorch, Keras backend with
TensorFlow support etc. We have created a simple GUI for
better user interaction, in which the user is able to navigate
tabs such as Browsing the Input Image file of any individual
from the dataset and segment it to obtain the tumor result,
accordingly, the segmented tumor in 2D can be converted
into a 3D structurewith the help of theothertabmentioned7
below and finally, by clicking the last tab that says “Show
Result” gives the output of a 3D Tumor which can be seen
from a Bird’s-Eye view.
6.1.2.2 OUTPUTS AND PERFORMANCE SCORES
After successfully evaluating the 3D model the output
obtained from input 3D MRI are also represented in 3D
format. Based on the intermediate snapshots generated for
different planes after the evaluation of the model, the most
stable plane is selected to display the tumor segmentation
which is shown in Fig. 5.

Fig. 5 – Tumor Segmentation
To represent the generated tumor segmentation Fig. 5, in 3D
format we convert the stored information about gradual
planes and display it in Bird’s-Eye view form. Fig. 7 clearly
depicts the desired output of an isolated tumor region in 3D.
The 3D output generated can be stored on the local device in
GIF or MP4 format to further utilize it for any biomedical
study based on brain tumor segmentation.
After successfully training the model certain parameters are
generated which are also known as performance scores.
Some of the parameters with high significance are Dice
coefficient (DSC), Recognition Rate (Sensitivity/Recall),
Specificity and Accuracy.
DSC also known as F1 Score is 0.579 along with the
Recognition Rate of 0.9106 and Specificity as 0.9857 for our
proposed system. After the complete evaluation the ACC
attained is 98.48%
…Eq. 10
…Eq. 11
…Eq. 12
…Eq. 13
Fig. 6 – Accuracy Graph
The Fig. 6 gives us visual representation of Accuracy varying
over subsequent epochs.
Fig. 7 – 3D Output
6.1.2.3 CLASSIFICATION METRICS
A Confusion matrix which is also known as error matrix is a
table that is often used to describe the performance of a
classification model, here, theclassificationisbetweentumor
pixels and non-tumor pixels classes. Confusion matrix 8 is a
special type of contingency table that contains two
dimensions actual and predicted values which gives an
identical set of classes known as TP (TruePositive),FP(False
Positive), TN (True Negative)and FN (False Negative).These
values summarize the results of our classification and can
also be used to derive other performance parameters either
through manual calculation or fetching the resultstoanother
model.

Fig. 8 – Confusion Matrix
Fig. 8 represents the values of identical set of classes
generated for our system.
TP = 601 FP = 813
FN = 59 TN = 56187
Using these values, we have calculated other performance
parameters such as NPV (Negative Predictive Value), FNR
(False Negative Rate), FPR (False Positive Rate), FDR (False
Discovery Rate), FOR (False Omission rate), MCC (Matthews
Correlation Coefficient).
…Eq. 14
…Eq. 15
…Eq. 16
…Eq. 17
…Eq. 18
As a result, the performance parametersNPV,FNR,FPR,FDR,
FOR, MCC obtained with the help of set of classes generated
for the system are 0.9989, 0.0893, 0.0142, 0.5749, 0.0010,
0.6164 respectively.
7. CONCLUSION
In this paper, we have successfully implemented U-Net
Architecture for precise and accurate brain tumor
segmentation. The algorithm takes various training images
provided in BraTS 2020 dataset and uses its predefined
convolutional patterns to determine the tumor region
effectively. Different mathematical models and strategies
such as Edgebaseddetectionmethod(EBD),Roughset-based
Fuzzy Clustering, Cross Entropy Loss are applied in an
abstract way for calculated process flow and operations. We
aim to accurately extract key features and characteristics for
the brain tumor by proposing this multimodal brain tumor
detection and segmentation system. By this system, we
developed a model to represent the whole overview of the
brain tumor region. Experimental results prove that our
system has a high ACC of 99.10% in case of2Devaluationand
98.48% for 3D evaluation.
8. FUTURE SCOPE
With regards to the future scope of our proposed system, we
aim to obtain the severity of Brain Tumor along withthetype
of Tumor. Recognition of Survival Rates will add value to the
system as well as for immediate treatment of a patient.
Detecting the growth pattern will help to curb the spread of
Brain Tumor.
9. REFERENCES
[1] W. M. J. X. LING TAN, "Multimodal Magnetic Resonance
Image Brain Tumor Segmentation Based on ACU-Net
Network," IEEE, vol. 9, pp. 14608-14618, Jan 2021.
[2] S. O. G. A. W. Mahnoor Ali, "Brain Tumor Image
Segmentation Using Deep Networks," IEEE, vol. 8, pp.
153589- 153598, Aug 2020.
[3] A. h. Abdel-Gawad, "Optimized Edge Detection
Technique for Brain Tumor Detection in MR Images,"
IEEE, vol. 8, pp. 136243-136259, Jul 2020.
[4] P. S. Nagwa M. Abouelenein, "HTT UNET: Hybrid Two
Track U-Net for automatic brain tumor segmentation.,"
IEEE, vol. 8, pp. 101406-101415, May 2020.
[5] Y. D. M. Z. Dongyuan Wu,"MultiFeaturesRefinementand
Aggregation for Medical Brain Segmentation," IEEE,vol.
8, pp. 57483-57496, Mar 2020.
[6] Y. H. a. Z. Zhang, "Deep learning assisted image
interactive framework for brain image segmentation,"
IEEE, vol. 8, Jun 2020.
[7] H. Guohua Cheng, "Adversarial Perturbation on MRI
Modalities in Brain Tumor Segmentation," IEEE, vol. 8,
pp. 206009-206015, Oct 2020.

[8] J. D. Jianxin Zhang, "Attention Gate Res U-Net for
Automatic MRI Brain Tumor Segmentation," IEEE,vol.8,
pp. 58533- 58545, Mar 2020.
[9] F. B. ,. Z. L. WEIGUANG WANG, "Learning Methods of
Convolutional Neural Network Combined With Image
Feature Extraction in BrainTumorDetection.," IEEE,vol.
8, pp. 152659- 152668, Aug 2020.
[10] Q. D. Q. W. PING LIU, "An EncoderDecoder Neural
Network With 3D Squeeze-and-Excitation and Deep
Supervision forBrainTumor Segmentation," IEEE,vol.8,
pp. 34029- 34037, Feb 2020.
[11] Q. S. W. WU DENG, "Deep LearningBased HCNN and
CRF-RRNN Model for Brain TumorSegmentation,"IEEE,
vol. 8, pp. 26665-26675, Jan 2020.
[12] S. R. TAMJID IMTIAZ, "Automated Brain Tumor
Segmentation Based on Multi-Planar Superpixel Level
Features Extracted From 3D MR Images," IEEE, vol. 8,
pp. 25335-25349, Dec 2019.
[13] L. R. R. A. CHANGHEE HAN, "Combining Noise-to-Image
and Imageto-Image GANs: Brain MR Image
Augmentation for Tumor Detection," IEEE, vol. 7, pp.
156966-156977, Sep 2019.
[14] M. M. H. ABDU GUMAEI, "A Hybrid Feature Extraction
Method With RegularizedExtremeLearningMachinefor
Brain Tumor Classification," IEEE, vol. 7, pp. 36266-
36273, Mar 2019.
[15] Q. G. Kai Hu, "Brain Tumor Segmentation Using Multi-
Cascaded Convolutional Neural Networks and
Conditional Random Field," IEEE, vol. 7, pp. 92615-
92629, Jul 2019.
[16] W. A.-A. Nancy M. Salem, "MultiClassification of Brain
Tumor Images Using DeepNeural Network," IEEE,vol.7,
pp. 69215-69225, May 2019.
[17] X. Y. AIMIN YANG, "Research on Feature Extraction of
Tumor Image Based on Convolutional Neural Network,"
IEEE, vol. 7, pp. 24204-24213, Feb 2019.
[18] G. M. P. M. Shaker, "Machine Learning Approach-Based
Gamma Distribution for Brain Tumor Detection and
Data Sample Imbalance Analysis," IEEE, Nov 2019.
[19] Z. G. Q. W. Qingneng Li, "Glioma Segmentation with a
Unified Algorithm in Multimodal MRIImages," IEEE,vol.
6, pp. 9543-9553, Feb 2018.
[20] W. L. ,. A. Z. Guotai Wang, "Interactive Medical Image
Segmentation Using Deep Learning With Image-Specific
Fine Tuning," IEEE, vol. 7, pp. 1562-1573, Jan 2018.

Brain Tumor Detection and Segmentation using UNET

More Related Content

Similar to Brain Tumor Detection and Segmentation using UNET (20)

More from IRJET Journal (20)

Recently uploaded (20)

Brain Tumor Detection and Segmentation using UNET