Brain Tumor Detection Using Convolutional Neural

Network
*Note: Sub-titles are not captured in Xplore and should not be used

line 1: 1st Given Name Surname line 1: 2nd Given Name Surname line 1: 3rd Given Name Surname
line 2: dept. name of organization line 2: dept. name of organization line 2: dept. name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 3: name of organization line 3: name of organization line 3: name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 4: City, Country line 4: City, Country line 4: City, Country
line 5: email address or ORCID line 5: email address or ORCID line 5: email address or ORCID

line 1: 4th Given Name Surname line 1: 5th Given Name Surname line 1: 6th Given Name Surname
line 2: dept. name of organization line 2: dept. name of organization line 2: dept. name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 3: name of organization line 3: name of organization line 3: name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 4: City, Country line 4: City, Country line 4: City, Country
line 5: email address or ORCID line 5: email address or ORCID line 5: email address or ORCID

Abstract—Brain tumor detection is a critical problem in benign tumors might have catastrophic implications if left
medical imaging, with the potential to have a significant impact untreated because of the organ's crucial role. Prompt medical
on patient outcomes. An overview of the difficulties, most action, such as surgery, radiation therapy, or chemotherapy,
recent developments, and current research in brain tumor is contingent upon early identification.
detection is given in this abstract. Medical imaging modalities
including Computed Tomography (CT) and Magnetic Medical imaging, such as magnetic resonance imaging
Resonance Imaging (MRI) are crucial for early diagnosis. The (MRI) and computed tomography (CT) scans, is critical in
assessment of radiologists was done manually in the past, but determining the presence of brain tumors. These imaging
with the introduction of deep learning techniques, especially technologies produce comprehensive cross-sectional pictures
Convolutional Neural Networks (CNNs), the area has of the brain, allowing doctors to see and describe aberrant
undergone a revolution. These AI-based methods have shown growths. However, deciphering these pictures can be a
encouraging results in automating the identification of brain difficult and time-consuming operation. It frequently need
tumors, increasing accuracy, and lessening the workload for the skills of highly trained radiologists who attentively study
medical personnel when combined with sizable and varied the images for minor indicators of malignancies. The
datasets. Yet, there are still issues, such as the requirement for expanding volume of medical imaging data, along with the
more datasets, the interpretability of AI models, and CNN
requirement for accurate and rapid diagnosis, has fueled the
achieved a really impressive accuracy of 97.87% in our study.
This paper's primary goal is to differentiate between normal
development of automated systems to assist healthcare
and aberrant pixels using statistical and texture-based criteria. workers.
CNNs, a subtype of artificial neural networks, have emerged
Keywords— CNN , Medical image , AI . as an effective tool for processing and evaluating medical
I. INTRODUCTION pictures. These networks are inspired by the human visual
system and excel in image identification and pattern
Brain tumors are a serious and frequently life-threatening recognition tasks. CNNs have demonstrated great
medical illness that affects people of all ages. The early and performance in a range of medical imaging applications,
precise diagnosis of brain tumors is critical because it can including tumor identification and classification.
have a significant influence on treatment success and patient
prognosis. Modern medical imaging technology, as well as The use of CNNs to detect brain tumors has several
advances in artificial intelligence, have opened the way for significant advantages. This is especially important in
novel methods to brain tumor identification. In this paper, we situations where quick decisions can save lives. CNNs can
will look at the use of Convolutional Neural Networks also help to reduce the risk of human error. In this paper, we
(CNNs) in the context of brain tumor detection, analyzing the proposed an efficient and skillful method which helps in the
importance of early detection, the problems faced by brain segmentation and detection of the brain tumor without any
tumors, and the possible benefits of using CNNs in the human assistance based on both traditional classifiers and
healthcare industry. Convolutional Neural Network.
Brain tumors are abnormal tissue growths that can be either II. LITERATURE REVIEW
benign or malignant. Depending on their size, location, and The most difficult and demanding tasks are those that
kind, these unregulated cell divisions can result in a variety involve separating the region of interest from an item, and
of neurological symptoms. Headaches, seizures, cognitive
the ambitious one that involves separating the tumor from an
deficits, and motor dysfunction are possible symptoms. Even
MRI brain picture[1]. International researchers are

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

researching in this area to obtain the best-segmented return
on investment. These days, segmentation using neural
networks produces notable results, and the use of this model
is growing daily[2]. Medical research on brain tumor
detection is essential as early identification is key to
effective therapy and a good prognosis for patients.

A. CNNs in Medical Imaging

Medical image analysis has been transformed by CNNs.

These neural networks are especially well-suited for Fig. 1
processing medical pictures since they are inspired by the
human visual system and perform very well in tasks
involving pattern identification[3]. CNNs' hierarchical layers Fig.1 depicts the approach of the current investigation. The
enable them to extract complicated information from current study's major processes include selecting a brain
pictures, which is a crucial capability for the identification tumor dataset, pre-processing MRI images, feature
of brain tumors. extraction, and classification using several classifiers.
B. Brain Tumor Detection and Classification using Deep
Learning.
A. DATASET
In deep learning, brain tumor identification and The image-based dataset used consisted of 3264 T1-
classification, Convolutional Neural Networks (CNNs) are weighted contrast-enhanced MRI scans [6]. This
trained on medical pictures, most often MRI scans. dataset had four categories of photos: glioma (926
images), meningioma (937 images), pituitary gland
Validation, model training, and data preparation are
tumor (901 images), and healthy brain (500 images).
crucial[4]. Tumor localization and classification are handled
The photos were all taken in the sagittal, axial, and
by the model, whose performance is assessed using coronal planes. Fig.2 shows instances of different
measures such as accuracy and F1-score. Accuracy and sorts of tumors and planes. A red outline has been
applicability are improving due to continuing research, with drawn around the tumor portion. The amount of
clinical integration and ethical issues being crucial. photos required for each subject varies.

C. Brain Tumor Segmentation with Deep Neural Networks.

The article "Brain Tumor Segmentation with Deep Neural

Networks" focuses on the accurate detection and delineation
of brain tumors in medical pictures, usually MRI scans, by
applying deep learning techniques, such as Convolutional
Neural Networks (CNNs)[5]. Network architecture design
and data preparation are essential. This method improves
tumor localization and supports the planning and tracking of
therapy.

III. PROPOSED METHODOLOGY

There are two unique models in our suggested technique for
segmentation and identification of brain tumors. The first
model utilized FCM to segment the tumor and typical
machine learning techniques to classify it, whereas the
second employed deep learning to detect tumors. For noisy
clustered data sets, FCM segmentation produces superior Fig. 2
results. Though it takes longer to execute, it retains more
information. B. Proposed model architecture
Convolutional layers for feature extraction are followed
by fully linked layers for classification in a suggested CNN
architecture for brain tumor detection. It combines transfer
learning with a pre-trained model to improve accuracy in
Fig.3.
 (a). Input Image (b). Skull Stripping (c). Filtering

(d).Image Enhancement (e). Segmentation (f). Tumor contouring

Fig. 5
Fig. 3
B. Performance Measures
C. Proposed Methodology for tumor detection using 5-
Layer Convolutional Neural Network[Fig.4] TP+TN
 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
TP+ FP+TN + FN
For tumor identification, the suggested technique leverages a
5-layer Convolutional Neural Network (CNN). It starts with TP
data gathering and preprocessing and then moves on to  Sensitivity =
TP+ FN
model architecture design. Model performance is assessed
throughout the training, validation, and testing phases, and
post-processing approaches improve outcomes. Throughout TP
 Precision(PPV)=
the procedure, ethical issues and clinical integration are TP+ FN
stressed.
TP
 Specificity=TNR=
Convolution Max pooling Flatten Dense Dense
TP+ FN
62*62*32 31*31*32 30752 128 1

 Hyper-parameters setting for CNN Model

TABLE.1 All the hyper-parameters value are

constituted in Table-I. Approximately 97.87% is
achieved as the accuracy.

STAGE HYPER-PARAMETER VALUE

Bias Zeros
INITIALIZATIO Glorot
Weight
N uniform
Learning table 0.001
Beta_1 0.9
Fig. 4
Beta_2 0.999

TRAINING Epsilon None

IV. EXPERIMENTAL RESULTS
In this section, we will discuss the experimental result of the Decay 0.0
proposed methodology. We divide this section into three
subsection which are : Amsgrad False

Epoch 10
A. Segmentation of the Brain Tumor using FCM

We segmented the tumor using our suggested methods

without losing any delicate information. We eliminated the
skull since the skull's participation in tumor segmentation is
essentially nil and uncertain in this process[Fig.5].
Classifiers Accuracy Recall Specifi Precisi Dice Jaccard
city on Score Index
1. Skull Stripping K-Nearest 89.39 0.94 0. 941
0.428 0.933 0.889
2. Filtering & Enhancement Neighbor 9
3. Segmentation using FCM Logistic 87.88 0.94 0.286 0.918 0.933 0.875
4. Tumor Contouring Regression 9
Multilayer 89.39 1.00 0 0.894 0.944 0.894
Perception 0
Naïve 78.79 0.79 0.714 0.959 0.870 0.770
Bayes 7
Random 89.39 0.98 0.167 0.903 0.943 0.892
Forest 3
SVM 92.42 0.98 0.428 0.935 0.959 0.921
3
C. Classification using Traditional Machine Learning
Algorithm
Classification using classic machine learning techniques
entails categorizing data points based on extracted features
and model training, laying the groundwork for a variety of
applications such as medical diagnosis and sentiment
analysis.

TABLE .2

D. Classification Using CNN

Convolutional Neural Networks (CNNs) classification is a
deep learning method that automatically assigns data points Fig. 6
to predetermined categories. CNNs excel at image and
geographic data tasks, extracting hierarchical characteristics Fig.6 displays our model's training and validation accuracy.
and allowing them to be used for image identification, The Keras callbacks function computed it. We measured
medical diagnosis, and natural language processing. training and validation accuracy using different numbers of
epochs. We discovered that after 9 epochs, the model had
The suggested five-layer approach yields a noteworthy the highest accuracy in both training and validation.
outcome for tumor identification. The suggested five layer
CNN model includes convolution, max pooling, flattening,
and two dense layers. Because CNN is translation invariant, LIMITATIONS AND FUTURE GOALS
data augmentation was performed before to training the
model. We measure the performance in two ways dependent Although our suggested models produced promising
on how the dataset is divided. We get 92.98% accuracy with classification results, there are still a number of concerns
a 70:30 splitting ratio, whereas training accuracy is 99.01%. that need to be addressed in future research. One of the
Then, in the second iteration, 80% of the photos were major challenges in employing deep learning-based
assigned for training and the remaining images were automated identification of brain tumors is the necessity for
accredited for testing, yielding 97.87% accuracy and 98.47% a large number of annotated pictures acquired by a skilled
training accuracy. So, when the divide is 80:20, our physician or radiologist. A huge dataset would be required
suggested model produces the best results. TABLE.3 to create a viable deep learning model. To the best of our
Represents the performance of the suggested CNN-based knowledge, the bulk of modern machine learning methods
approach. for medical imaging suffer from this limitation. Although
the bulk of previous research are presently making their
We achieved 97.87% accuracy using five-layer CNN, which datasets public in an effort to remedy this issue.
is impressive. We tested with a variety of layer counts, but Nonetheless, the amount of adequately and accurately
the differences in results were not statistically significant documented data remains quite restricted.
when employing our five-layer CNN model. When we
increased the number of layers, we saw an increase in Adopting zero-shot, few-shot, and deep reinforcement
computing time, the complexity of the technique batch size, learning (DRL) approaches in the future might help us
and steps per second. Furthermore, we set the dropout value overcome this challenge. Zero-shot learning may create a
to 0.2 but did not calibrate the model since the accuracy recognition model using unseen test samples that have not
flattened[7]. As a consequence, this model has the highest been tagged for training. As a result, zero-shot learning can
accuracy without the use of dropout. handle the problem of tumor classes without training data.
Furthermore, utilizing the few-shot learning approach, a
TABLE.3 Performance of the proposed CNN model deep learning model may learn information from a limited
number of labeled cases per class. DRL, on the other hand,
No
Training Testing Splitting Accuracy has the potential to lessen the requirement for exact
Image Image Ratio (%)
annotations and high-quality photos.
1 152 65 70 : 30 92.98

2 174 43 80 : 20 97.87 Another limitation of this study is that, while the suggested
technique performed well on two publically available
datasets, the work has not been verified on actual clinical
trials. This is also true for practically all of the models
examined in this study. When actual clinical data becomes
available, we intend to test our model on it. We may directly
compare the performance of our suggested models to
experimental procedures in this manner. Another potential
path for the future is to utilize more layers or other
regularization techniques to work with a tiny picture dataset
using a CNN model.
REFERENCES [5] Havaei, M., Davy, A., Warde-Farley, D., et al. (2017). Brain Tumor
Segmentation with Deep Neural Networks.
[6] Lee JY, Lee KS, Seo BK, Cho KR, Woo OH, Song SE, et al.
[1] Havaei, M., Davy, A., Warde-Farley, D., et al. (2017). Brain Tumor Radiomic machine learning for predicting prognostic biomarkers and
Segmentation with Deep Neural Networks. molecular subtypes of breast cancer using tumor heterogeneity and
[2] Kamnitsas, K., Ledig, C., Newcombe, V. F., et al. (2017). Efficient angiogenesis properties on MRI. Eur Radiol.
multi-scale 3D CNN with fully connected CRF for accurate brain [7] Zhou, T., Basu, S., & Chakraborty, D. P. (2019). Brain
lesion segmentation. Tumor Detection and Classification Using Deep
[3] Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A., Ciompi, F., Learning and Improved Radiomics.
Ghafoorian, M., ... & Sánchez, C. I. (2017). A survey on deep learning
in medical image analysis. Medical Image Analysis, 42, 60-88.
[4] Mehdipour, S., Mohseni, H. R., & Far, B. H. (2018). Brain Tumor
Detection and Classification using Deep Learning.