DIRECTIONAL CLASSIFICATION OF BRAIN TUMOR IMAGES FROM MRI USING CNN-BASED DEEP LEARNING
- 1. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 11 Issue: 01 | Jan 2024 www.irjet.net p-ISSN: 2395-0072
© 2024, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 342
DIRECTIONAL CLASSIFICATION OF BRAIN TUMOR IMAGES FROM MRI
USING CNN-BASED DEEP LEARNING
G MAHESH CHALLARI 1, J PRASHANTHI 2
1 Assistant Professor in Department of cse at Sree Dattha Institute of Engineering And Science,
2 Assistant Professor in Department of cse at Sree Dattha Group of Institutions.,
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - A brain tumour is a dangerous development of
unnatural cells in the brain. If not treated in time, it might
prove fatal. It is therefore crucial to find the tumor early and
start therapy as soon as possible. People with brain tumours
had a significant fatality rate before the discovery of early
diagnosis. The mortality rate lowers considerably after an
early diagnosis is established. Accurate early diagnosis of a
brain tumour increases a patient's chance of survival. The
customary system used to detect tumors involved physicians
studying the MRI scans and analyzing the abnormalities in the
image. However, with the increase in the size of data and
limited amount of time it becomes extremely strenuousforthe
physicians to analyze the image. Our research has resulted in
the release of a convolutional neural network model for the
detection of brain tumours, whichfurtherclassifiesthetumour
into glioma, pituitary, or meningioma. This automated model
is improving the detection and classification accuracy of
tumours, demonstrating that it is a useful tool for physicians.
The same brain MRI database is used to train and test all the
types under consideration, including CNN, ResNet50,
MobileNetV2, and VGG19. The effectiveness of each type is
reviewed. Accuracy, error rate, and time to train are just few
of the criteria used to evaluate the results from each CNN
variant.
Key Words: Convolutional Neural Networks, Brain Tumor,
MRI, Medical Disorder, Healthcare.
1. INTRODUCTION
Analyzing biomedical data is growing significantly and is
crucial in diagnosing and properly treating disease. Brain
tumor contains more complex image data that needs image
processing to analyze. There is a higher death rate and
improper treatment of brain tumors, as the statistics taken
by the National Brain Tumor Foundation (NBTF)worldwide
[1]. Several approaches (or) frameworks have been
developed to consider brain tumors in recentyears.Itcomes
across data classification, proper treatment planning and
predicting outcomes. Images of tumours in the brain's
structure are difficult to classify because of issues with
contrast, noise,andmissing boundaries.Magnetic Resonance
(MR) imaging [2, 3], Positron Emission Tomography (PET)
[3, 4], and Computed Tomography (CT) scan [5, 6] are used
to assess the efficacy of the diagnostic procedure by
analysing the aforementioned elements. Scannersuseimage
processing to look for signs of illness. The scanning process
is done to diagnose and treat the brain tumours better,
which identifies tumor images to be analyzed effectively.
Suppose the identification of predicting the tumor cells and
locating them helps to diagnosis the disease properly.
Making the analyses more qualitative and amount quantity
improves the characteristics of a tumor diagnosis through
data segmentation. Based on the development of the MR
images generation, it can be processed manually, semi-
automatically, and fully automatically [5]. Based on the
image processing related to the medical field, information
should be accurate while processing the image based on the
segmentation and classification process. While processing
the images, the execution should be time-consuming [6].
2. LITERATURE REVIEW
In computer vision, image segmentation is a highly effective
technological procedure. In the field of image processing,
segmentation is a useful tool. Pixels are organized into
groups called segments, which are themselves artefacts.
When an image is segmented, it is no longer necessary to
analyze the entire thing at once. The three steps involved in
the processing of an image are depicted in overall diagnosis
of "skull," "brain," or "tumour" might be possible by
labelling. One application of object detection is to identify
certain objects inside a picture. In order to properly identify
and categorize objects, segmentation is essential
In this section, we take a look at the various segmentation
methods currently available in the literature for MR image
processing. The author [66] proposes an appropriate, novel
approach to tissue segmentation from MRI brain imaging.
The "WM and GM and CSF" segmentation is useful for
studying diseasesanddesigningtreatments.Toeliminatethe
graininess and sharpen the image, anisotropic diffusion
filtering is used. Tests of the proposed technique on 10 MRI
images have been conducted, and the results have been
compared to those obtained using existing methods
(including "Otsu MT" "fuzzy C-means". The results of the
experiments show that, in comparison to the existing
approaches, the average segmentation accuracyisimproved
by 96.79%, the specificity is 96.55%, and the sensitivity is
96.55%
Brain tumour identification and increased breast cancer
detection in MRI images were both targeted by the author of
[82], who proposed a template-basedKmeansandenhanced
- 2. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 11 Issue: 01 | Jan 2024 www.irjet.net p-ISSN: 2395-0072
© 2024, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 343
fuzzy C means (TKFCM) system. This algorithmoutperforms
the competition even when presented with a noisy MR
image. In comparison to thresholding, area expanding,
region splitting and merging, ANN, TK-means, and FCM, this
algorithm performs better. It used a simulator to test out its
proposed solution and see how well it worked. Besides, it
has proven a consensus reached to be more reliablethanthe
conventional statistical classification results. However, the
time required for the TKFCM algorithm to detect brain
tumors is concise compared to other traditional methods.
3. SYSTEM ANALYSIS
3.1.1 DETECTION OF BRAIN TUMOR USING
RESUNET ARCHITECTURE
Using a classification technique, a brain tumour detection
system aids in diagnosing and treating patients recently
diagnosed with a tumour in the brain. Early diagnosis and
treatment can be provided by using the MRI classification
method, which helps identify the brain's tumour and
determine the tumor’s density. U-Net architecture is
associated with a variety of classification methods in the
literature. Based on a comparison oflow-level andhigh-level
feature information, this work proposes Residual U-NET
with improved local feature information for improved
medical image segmentation. As part of the modified
Residual U-Net model, the dropout and wide context layers
are addressed in addition to the residual module and
attention gate. Large sensitive scaled information and
small-scale images benefit from adding salient feature
information. After making certain adjustments, the gate
attention model performed better than state-of-the-art
methods like U-Net and CNN Densely.
3.1.2 LGG Dataset
This dataset has MR images of the brain and manual flare
abnormality segmentation masks. In total, the Cancer
Genome Atlas (TCGA) collection contains 3929 images,
including 2556 non tumor images and 1373 cancer images
from 110 individuals.
Figure 3.1.2 Brain Tumor image and Mask image
The ratio of training and test images are tested 70:30. The
size of all images in this dataset is 256 x 256 pixels. Access
the dataset at 'https: //www.kaggle.com/mateuszbuda/lgg-
mri-segmentation'...
4. PROPOSED METHODOLOGIES
4.1.1 Image Pre-processing
The deep learning model is crucial in making the network
less sensitive to the noise in the data and imagesitisgiven to
analyze. Because of how consistently it analyses images
across the board, the N4ITK method isutilizedforcorrecting
bias-based image processing. Several algorithms exist that
perform the data pre-processing for brain tumor. From the
existing N4ITK algorithm provides reliable dataasitcollects
the data based on the capability of bias field for MRI Images.
4.1.2 Modified Residual U-Net
As residual networkassociate the U-Netbackbonewithmore
residual blocks in Deep framework to gradient mitigation
problem. In this U-Netarchitecture, the convolutional neural
network is related to performing image segmentation into
three elements, Encoderanddecoder,basedontheprocessof
contraction and expansion. For the purpose of expandingthe
blocks, the suggested architecture takes into account the
Convolutional 2D, Max pooling 2D, Wide Context, Transpose
2D, Residual Block, and Concatenate metrics.
4.1.3 Max Pooling Layer
Layer will supply the down feature extraction mapping,
highlighting the feature patches depending on the feature
map approach, in addition to max and average pooling.
Maximum presence pooling uses extracted features to
calculate an average presence and then uses that to activate
the presence ofthosefeatures.Simplysaid,thisconvolutional
layer is the neural network-based convolutional layer that is
used to generate the mapping feature extraction by filtering
learning based on input images. It is determined through
research that the layer functions best when extracting
simple features.
4.1.4 Dropout Layer
It is the method used to stay away from the convolutional
neural network model's overfitting issue. When the training
phase is updated, the main cause of dropout is a random
setting on the hidden edge units being set to 0 as neurons.
Using a sigmoid activation function and a numerical
representation of the residual u-net
- 3. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 11 Issue: 01 | Jan 2024 www.irjet.net p-ISSN: 2395-0072
© 2024, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 344
4.1.5 Attention Gate with Residual U-Net
The coming step of data preprocessing is to handle missing
Two parallel connections are used as input for the two
convolutional layers in a broad context block. Use both 1 x N
and N x 1 filters on the initial link. The connection uses 1 x N
filter with a convolutional model and N x 1 filter with a
convolutional model. In this process, the information’s are
extracted to form the data similarity byconsidering the wide
context. It is further classified into different convolutional
sub-class networks for brain tissue.
4.1.6 Wide Context Information
Two parallel connections are used as input for two
convolutional layers in a large contextblock. Use both 1 x N
and N x 1 filters on the initial link. The connection uses 1 x N
filter with a convolutional model and N x 1 filter with a
convolutional model. In this process, the information’s are
extracted to form the data similarity byconsidering the wide
context. It is further classified into different convolutional
sub-class networks for brain tissue.
Figure 4.1.3 Max pooling layer
Figure 4.1.4 Dropout layer
Figure 4.1.5 Residual U-Net Architecture with Gate
Attention Mechanism
5. RESULTS
The quantity of photos given careful consideration is one
indicator of performance in the analysis of results. Tensor
Flow, along with the proper modules, can be used as a
checkpoint and scheduler for the learning process. Patient,
RNSseq cluster, laterality, tumour location, and ethnicity are
justsomeof the granular characteristics thatmaybefoundin
the Kaggle dataset. Those datasets with photos are the result
of a checkpoint- driven sorting procedure. Patient id, picture
path, and mask path are all included in the final dataset as
shown in Table 3.1. To compare the count value to the mask
and find the mask count plot.
S.No. Mask Dtype
1 0 2556
2 1 1373
Table 5.1 Mask Count part
To locate the tumour on the mask, we use photos of brain
tumours with varying pixel colours. Using parameters like
image path, mask path,and mask,a new datasetisgenerated.
Brain_df_train, test size = 0.15 is used to test and divide
trained dataaccording on the chosen model. After that,those
pictures are confirmed
Using a classifier model trained with data from the Residual
Network ResNet50, we examine the photos. Conv2D, Input,
ZeroPadding2D, Batch Normalization, Activation,
MaxPooling2D, and batch normalizedimagesareclassifiedin
the input layer. Using an image classification model to
determine whether or not the tumour is present, the
classifier model achieved a loss in data of 0.2353, with an
accuracy of 94.745%. Tables 3.2 and 3.3 depict the metrics
used to generate the confusion matrix from the original and
forecast images: accuracy score, confusion matrix, and
classification report
In
p
ut
ou
tp
ut
1x9
Residu
al U-
Net
9x
1
Re
sid
ual
U-
Ne
t
1x
1
1
R
es
id
u
al
U
-
N
et
11
x1
Re
sid
ual
U-
Ne
t
1x
1
3
R
es
id
u
al
U
-
N
et
13x
1
Res
idu
al
U-
Ne
t
1x
1
5
R
es
id
u
al
U-
N
et
15x
1
Res
idu
al
U-
Net
- 4. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 11 Issue: 01 | Jan 2024 www.irjet.net p-ISSN: 2395-0072
© 2024, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 345
Table 5.2 Attribute metrics based on mask value
Table 5.3 Attribute metrics based on mask value
(Accuracy; Macro Avg; Weighted Avg)
Training and validation loss values, or how well they
predicted, are plotted against epochs (the X-axis in Figure
3.6). The loss ranges from zero to one. For the proposed Res
U-Net architecture, the loss is calculated using data from the
first 30 epochs. Since the validation set is probably not
representative of the complete dataset, the validation loss
varies greatly. The validation sample should be randomized
or resampled, in my opinion
Figure 5.1 Epochs Vs Training &Validation loss
Figure 5.2 depicts how the dice coefficient for the proposed
Residual U-Net architecture is calculated using epochs
ranging from 0 to 30 for both training and testing data. The
value of the dice rises steadily with the number of epochs
Accuracy for the suggested Residual U-Net architecture is
shown in Figure 3.8, where epoch ranges from 0 to 30 are
used for the trained data and Val. As more time has passed
(more epochs have been added), theaccuracyvaluehasrisen
steadily
Figure 5.3 Epochs Vs Dice coefficient
Data pre-processing, encoding, and decoding on MRI image
datasets are all carried out using the suggested architecture.
Segmentation of improved images is usedasthebasisforthis
analysis.Loss andAccuracyaredeterminedbycomparingthe
values of the training and test epochs, respectively. The data
categorization model uses the Epochs value to calculate data
loss and accuracy. The proposed Res U-Net achieves better
accuracy than the currentResU-Netmethodasthenumberof
iterations increases, as measured by the loss function, Dice
co-efficient value, 95% Accuracy.
Figure 5.4 Epochs Vs Accuracy
6. CONCLUSION
In the recent era, healthcareplays a significant role based on
biomedical datasets related to brain tumor images and other
classification and segmentation processes. Various types of
tumor are discussed based on different grades. At first, we
talk about how the dataset influences the preprocessing of
images. After that, a set of feature photos is generated and
categorizedusingestablishedmethods.Tumourdetectionand
image segmentation processing are described. To better
segment medical imagesandfocusontheattentionmodulein
tumour images, wefirstpresentthedesignofResidualU-NET
with some improved local feature information. In this, local
information on image feature are considered by improving
the image segmentation through the proposed Residual U-
Net as it integrates both residual and attention gate modules
and considers both dropout and wide context layer. Focal
Trversky Loss and Trversky Accuracy are computed using
training and test data, and Epochs value variation is
evaluated for large-scale images through whole, core, and
enhancing image segmentation.
- 5. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 11 Issue: 01 | Jan 2024 www.irjet.net p-ISSN: 2395-0072
© 2024, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 346
Brain tumor segmentation using Spatial Attention ResU-Net
gives best results compared to existing U-Net models. With
increasing training data, wecan furtherimprovediceandIoU
scores. And also with more research on the classes labelled
on the data we can predict the stage and severity of the
disease. After segmenting the MR Image, it is possible to
measure the tumour's size in brain tissues
REFERENCES
[1]. Robert Tufel, “The National Brain Tumor Foundation”:
Giving Help, Giving Hope”; Journal of Neoplasia, 3(3): 264-
265, 2001.
[2]. Soltaninejad, Mohammadreza, Zhang, Lei, Lambrou,
Tryphon, Yang,Guang, Allinson, Nigel and Ye, Xujiong,” MRI
Brain Tumor Segmentation and Patient Survival Prediction
Using Random Forests and Fully Convolutional Networks”,
ComputerMethodsandPrograms inBiomedicine, 157,69-84,
2018.
[3]. Langen KJ, Bartenstein P,Boecker H, Brust P, CoenenHH,
Drzezga A, Grünwald F, Krause BJ, Kuwert T, Sabri O, Tatsch
K, Weber WA, Schreckenberger M, “German guidelines for
brain tumour imaging by PET and SPECT using labelled
amino acids”, Nuclear Medicine, 50(4), 167-173, 2011.
[4]. Irene Neuner, Joachim B. Kaffanke, Karl-Josef Langen,
Elena Rota Kops, Lutz Tellmann, Gabriele Stoffels, Christoph
Weirich, Christian Filss, Jürgen Scheins, Hans Herzog & N. Jon
Shah, “Multimodal imaging utilising integrated MR-PET for
human brain tumour assessment”, European Radiology
volume 22, 2568–2580, 2012.
[5]. B. Johnston; M.S. Atkins; B. Mackiewich; M. Anderson,
“Segmentation of multiple sclerosis lesions in intensity
corrected multispectral MRI”, IEEE Transactions on Medical
Imaging, 15(2), 154-169, 1996.
[6].Erena Siyoum Biratu; Friedhelm Schwenker; Yehuala shet
Megersa Ayanoand TayeGirma Debelee, “A Survey of Brain
Tumor Segmentation and Classification Algorithms”, J.
Imaging, 7(9), 2021.
[7]. Lim Jia Qi, Norma Alias, FarhanaJohar, “Detection of brain
tumour in 2D MRI: implementation and critical review of
clustering-based image segmentationmethods”, Oncologyand
Radiotherapy, 1(52), 1-10, 2020.
[8]. Dr. Thejaswini P; Ms. Bhavya Bhat; Kushal Prakash,
“Detection and Classification of Tumour in Brain MRI”, I.J.
Engineering and Manufacturing, 2019.
[9]. G. Preethi; V. Sornagopal, “MRI image classification using
GLCM texture features”, International Conference on Green
Computing Communication and Electrical Engineering
(ICGCCEE), 2014.
[10]. Sindhu Devunooru, AbeerAlsadoon, P. W. C.
Chandana&Azam Beg, “Deep learning neural networks for
medical image segmentation of brain tumours for diagnosis”,
Journal of Ambient Intelligence and Humanized Computing,
12, 455-483, 2021.