Kathford Journal of Engineering and Management,

ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

Brain Tumor Detection Using Convolutional Neural

Networks (CNNs)
Sangam Aryal1, Sangeeta Sharma1, Siddhant Sedai1, Prashraya Aryal1, Jalauddin Mansur*,1
1
Department of Computer Engineering, Kathford Int’l College of Engineering and Management,
Institute of Engineering, Tribhuvan University, Kathmandu, Nepal

*Corresponding Author: [email protected]

ABSTRACTBrain tumors are a serious medical condition that requires early detection for
successful treatment. However, accurate diagnosis can be difficult and time-consuming, and
current methods such as MRI scans can be expensive and may require highly trained specialists to
interpret the results. A model of a brain tumor detection system using Convolutional Neural
Networks (CNNs) has been proposed to address these challenges. To use this model, a dataset of
medical images of the brain is collected, the dataset is then preprocessed, and the relevant feature
is extracted from the images using CNNs. The developed CNN model is designed and trained to
accurately detect the presence and location of brain tumors in the images. Optimization of the CNN
model's performance is done by experimenting with different architectures, hyperparameters, and
optimization techniques, and its performance is evaluated using metrics such as accuracy,
sensitivity, specificity, and F1 score. The model training was carried out on MRI images containing
tumors and without tumors. The developed CNN-based model achieved impressive accuracy in
detecting brain tumors, demonstrating high precision and recall rates. This brain tumor detection
system has the potential to significantly improve the accuracy and efficiency of brain tumor
diagnosis, leading to better treatment outcomes and reducing the burden on healthcare systems.

KEYWORDSMagnetic resonance imaging, Optimization techniques, Medical imaging,

Image Processing, Neural Network Architectures, Machine Learning

1. INTRODUCTION
have been successfully applied in medical
Brain tumors are a serious medical condition image analysis, including brain tumor
that can have significant impacts on a patient's detection, to improve the accuracy and
health and well-being. Early detection of brain efficiency of diagnosis. Additionally, this
tumors is critical for successful treatment, but
project has applications in research, education,
accurate diagnosis can be difficult and time-
and public health, making it a valuable area of
consuming (Abdusalomov et al., 2023).
study for medical professionals and computer
Current methods for brain tumor detection,
scientists alike. The main objectives were to
such as MRI scans, can be expensive and may
optimize the CNN model's performance by
require highly trained specialists to interpret
experimenting with different architectures,
the results. Convolutional Neural Networks
hyperparameters, and optimization techniques
(CNNs) are a type of deep learning model that
and evaluate the performance of the CNN
has shown promising results in image
model using metrics such as accuracy, loss,
recognition and classification tasks. CNNs

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

and F1 score(Johny et al., 2020). This research carefully crafted design that allows for
endeavors to bridge this gap by leveraging a increasing the depth and width of the network
unique dataset sourced directly from real while keeping the computational budget
hospital environments. The acquisition and constant. His results seem to yield solid
utilization of authentic patient data from Bir evidence that approximating the expected
Hospital present a significant opportunity to optimal sparse structure by readily available
enhance the robustness and clinical relevance dense building blocks is a viable method for
of our proposed brain tumor detection model. improving neural networks for computer
The utilization of this real-world dataset vision. In this study, a review of the previous
addresses the pressing need for models trained work of the last ten years is discussed for
on diverse, clinically representative cases, comparison purposes. CNN technique is used
enabling our approach to potentially excel in for the classification of the grey-scaled
practical medical settings. The utilization of segmented MR Images to get accurate results
hospital data not only contributes to the for treatment planning and improvement.
authenticity and richness of our research but This study provides help to radiologists,
also underscores our commitment to ethical doctors, and surgeons in the diagnosis of
practices and patient privacy. disease in a very short time and with high
accuracy. This study will contribute
2. LITERATURE REVIEW effectively to the field of image
processing(Szegedy et al., 2015).
Sankari et al. along with the other researchers
Mustaqeem et al. proposed a hybrid
came up with a model for cancer diagnosis for
segmentation technique including Watershed
a brain tumor which is the toughest task. Most
and Thresholding-based segmentation
research has been done in this field using
techniques. Firstly, the quality of the scanned
PCA, Route set theory, and the Wavelet
images is enhanced and then morphological
method. The authors here used Convolutional
operations are applied to detect the tumor
neural networks to solve the problem.
along with their proposed hybrid
Respected authors proposed image de-
segmentation. The proposed system is easy to
noising, intensity normalization, and bias-
execute and thus can be managed easily.
field correction methods for the image pre-
Obtained MRI images are displayed in two-
processing task. They used a bilateral filter to
dimensional matrices having pixels as their
remove the noise from the MRI images.
elements. Gaussian low pass filter and
Histogram Equalization is used for enhancing
averaging filters are used to remove salt and
and feature extraction of the image. And
pepper noise from the image. The filter pixel’s
finally, CNN is used to classify the
value is replaced with its neighborhood
images(Sankari et al., 2020).
values. Gaussian high pass filter improves the
Szegedy et al. proposed a deep convolutional
boundaries of the objects in the image.
neural network architecture codenamed
Threshold segmentation is used to convert the
Inception, which was responsible for setting
grayscale image into a binary image format.
the new state of the art for classification and
Watershed Segmentation is used to group
detection in the ImageNet Large-Scale Visual
pixels of an image based on their intensities.
Recognition Challenge 2014(ILSVRC14).
Morphological operators are applied to the
They used this architecture to improve the
converted binary image to separate the tumor
utilization of computing resources inside the
part of the image(Mustaqeem et al., 2012).
CNN network. This was achieved by a

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

Astina Minz et al. implemented an operative CNN Architectures, Dataset Characteristics

automatic classification approach for brain and Transfer Learning," featured in IEEE
images that projected the usage of the Transactions on Medical Imaging. The
AdaBoost gadget mastering algorithm. The authors meticulously examined various CNN
proposed system includes three main architectures, providing nuanced insights into
segments. Pre-processing has eliminated the their advantages and limitations within the
noises in the datasets and converted MRI context of medical imaging analysis. Beyond
images into grayscale. Median filtering and architecture, the review delved into the critical
thresholding segmentation are implemented in role of dataset characteristics, emphasizing
the pre-processed image(Minz & Mahobiya, considerations such as size, diversity, and
2017). annotation quality that are pivotal for training
Havaei et al. introduced an innovative robust CNN models. Furthermore, Shin and
approach to brain tumor segmentation colleagues explored transfer learning,
employing deep neural networks. Their work, discussing its potential benefits and
addressing challenges associated with
documented in "Brain Tumor Segmentation
adapting pre-trained CNNs to medical
with Deep Neural Networks" published in datasets. This comprehensive examination
Medical Image Analysis, signifies a equips researchers and practitioners with a
transformative contribution to medical image thorough understanding of the complexities
analysis. Leveraging Convolutional Neural involved in deploying CNNs effectively for
Networks (CNNs), the authors achieved computer-aided detection, influencing
remarkable accuracy in delineating tumor subsequent research endeavors aimed at
refining and optimizing CNN models for
boundaries from complex imaging data. This
heightened precision in medical imaging
method not only demonstrated a significant tasks(Shin et al., 2016).
improvement in segmentation precision but
also laid the groundwork for subsequent 3. METHODOLOGY
advancements in automated tumor analysis. 3.1. Algorithm Convolutional Neural
Beyond its immediate impact, the paper's Network (CNN)
insights have influenced the broader trajectory Convolutional Neural Networks (CNNs) are a
popular algorithm for image recognition and
of research in medical image
classification tasks, including the detection of
segmentation(Havaei et al., 2017). brain tumors from medical images such as
Shin et al. conducted an extensive literature Magnetic Resonance Imaging (MRI) scans
review on the application of Deep (Lamrani et al., 2022).
Convolutional Neural Networks (CNNs) in The following is the architecture of CNN used
computer-aided detection in their 2016 in our model:
publication, "Deep Convolutional Neural
Networks for Computer-Aided Detection:

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

Input Layer

Convolution Convolution
al layer Activation al layer
32 64
Kernel Size: Layer Kernel Size:
Feature Feature
33 Function: 33
Maps Maps
Filters: 32 relu Filters: 64

MRI Image
224  224 pixels RGB
Feature Extraction

Activation Convolution Activation

al layer 64 Max
Layer Feature Pooling Pool Layer
Kernel Size:
Function: Maps Size Function:
relu 33 relu
22
Filters: 64

Convolution Activation Max

Max al layer 128 Dropout
Pooling Pool Feature Layer Pooling Pool layer
Kernel Size:
Size Maps Function: Size
33 relu
22 22
Filters: 128
Classification

Is there Activation
Dropout Layer
a layer Function:
Flattening
tumor? relu

Figure 1. CNN Architecture

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

3.1.1 Input Layer prevent overfitting by dropping the fraction of

This is the first layer of our model in which neurons so that the model can learn more
gray-scaled MRI images of size 224x224 generalized features.
pixels from the training dataset are passed into 3.1.5 Output Layer
it. This is the final layer in our model which
3.1.2 Convolution Layer produces a single output of either 1 or 0 where
There are four convolution layers in our model 0 represents the presence of a brain tumor and
in which each layer has a kernel size of 3x3. 1 represents the absence of a brain tumor. The
The first convolution layer produces 32 activation function used in the output layer is
feature maps. The second layer and the third sigmoid as its value ranges from 0 to 1.
produces 64 and the final convolution layer
produces 128 feature maps. After each 3.2 Process Model
convolution layer, the ReLU activation
function is used to introduce non-linearities in This process model aligns with the typical
the model by setting all the negative values to steps in building a machine learning system:
zero. 1. Data Acquisition: This initial step
3.1.3 Pooling Layer involves collecting data from various
There are 3 Max pooling layers in our model sources relevant to the problem at

Data Acquisition

Data Processing

Model Development

Model Training

Model Evaluation

Development

Maintenance and Support

Figure 2. Process Model

hand(Seth, 2021). It could be
of size 2x2 which picks the maximum value of structured data from databases,
each 2x2 region in the feature map. unstructured data from text or images,
3.1.4 Dropout Layer or even data gathered through sensors
The dropout layer in our model is used to or IoT devices(Seth, 2021).

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

2. Data Preprocessing: Once the data is 6. Deployment: Once the model

collected, it needs to be cleaned and performs satisfactorily, it's deployed
preprocessed. This step involves into the production environment. This
handling missing values, dealing with phase involves integrating the model
outliers, normalizing or standardizing into the system where it will be used to
data, encoding categorical variables, make predictions or perform
and splitting the dataset into training, tasks(What Is the AI Life Cycle? -
validation, and test sets(“Data Data Science Process Alliance, n.d.).
Preprocessing in Data Mining,” 2019).
7. Maintenance & Support: Even after
3. Model Development: With deployment, the model requires
preprocessed data, the model ongoing maintenance. This includes
development phase begins. This monitoring its performance, retraining
involves selecting an appropriate the model periodically with new data
machine-learning algorithm or model to prevent model degradation, and
architecture based on the nature of the addressing any issues that arise during
problem (classification, regression, its operational use. Additionally,
clustering, etc.)(Machine Learning providing support for the end-users
Model Development and Model and ensuring the model's continued
Operations, n.d.). relevance and effectiveness is
crucial(Seth, 2021).
4. Model Training: The selected model Using this process model can help
is trained using the training dataset. ensure that the brain tumor detection
During training, the model learns system is developed in a structured and
patterns and relationships within the efficient manner, while also taking
data to make predictions or into account the needs of stakeholders
classifications. (Supervised Learning, and the clinical setting.
2023)
3.3 Dataset Distribution
5. Model Evaluation: After training, the
model's performance needs
evaluation. It's tested on the validation
set to assess its accuracy, precision,
recall, F1 score, or other relevant
metrics. This step helps in tuning Figure 3. MRI scans of tumorous brain
hyperparameters or adjusting the
model architecture to improve
performance(Evaluation Metrics | 12
Must-Know ML Model Evaluation
Metrics, n.d.). Figure 4. MRI scans of non-tumorous
brain

images consisting of 3734 samples containing

The model training was carried out on 7295 tumors and 3561 samples containing no

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

tumors. The data is further divided into 80% layer, four convolutional layers, two dropout
of the data as data training, 10% of the data as layers, and a fully connected(dense) layer;
data validation, and 10% of the data as data each experiment uses 15 epochs and 32
testing. The data is run 15 times, each using batches.
the CNN model which consists of one input

Figure 5. Dataset Distribution

3.4 Block Diagram testing set is used to evaluate the final

performance of the model.
1. Labeled Dataset: The process begins
3. Set Hyperparameters: Before
with a labeled dataset of brain tumor
building the CNN architecture, the
images. Each image is labeled with the
hyperparameters of the model need to
type of tumor it contains (e.g., glioma,
be set. These hyperparameters include
meningioma).
the number of layers in the network,
2. Data Partitioning: The dataset is then
the number of neurons in each layer,
split into three parts: training,
and the learning rate.
validation, and testing. The training set
is used to train the CNN model, the
validation set is used to fine-tune the
model's hyperparameters, and the

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

Labelled
Dataset Data
Partitioning

Training and
Validation Testing

Build Model
Final Performa-
CNN Training
Set hyper- nce
architecture Model Computa-
parameters
tion

Figure 6. Block Diagram

metrics help to assess how well the
4. Build CNN Architecture: The next model can correctly classify brain
step is to build the architecture of the tumors(Machine Learning Model
CNN. This involves defining the Development and Model Operations,
number and type of layers in the n.d.).
network, as well as the connections
between them. 7. Final Model: Based on the
performance evaluation on the
5. Model Training: Once the validation set, the model may be fine-
architecture is defined, the CNN tuned by adjusting its
model is trained on the training set. hyperparameters. Once the model is
During training, the model is finalized, its performance is evaluated
repeatedly exposed to the training on the testing set. The performance on
images and their labels. The model the testing set is the final measure of
learns to identify features in the how well the model will generalize to
images that are associated with new data(Machine Learning Model
different types of tumors. Development and Model Operations,
n.d.).
6. Performance Computation: After
training, the performance of the model 3.5 Performance Evaluation Metrics
is evaluated on the validation set. This
involves calculating metrics such as 3.5.1 Precision
accuracy, precision, and recall. These Also known as Positive Predictive Value,

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

precision measures the proportion of true The harmonic mean of precision and recall,
positive predictions (correctly identified providing a balanced measure that considers
tumors) among all the samples predicted as both false positives and false negatives. It's
tumors. It helps in understanding the model's particularly useful when dealing with
ability to avoid false positives. imbalanced datasets.
Precision = (TP+FP)/TP F1score = (2*precision*recall)/(precision +
recall)
3.5.2 Recall
This metric indicates the proportion of true 4. RESULTS AND DISCUSSIONS
positive predictions (correctly identified
tumors) among all actual tumor samples. It The results of the training are shown below.
measures the model's ability to capture all This model has an accuracy of 83.17% and a
positive instances and avoid false negatives. loss of 0.3614 on test data, an accuracy of
Recall= (TP)/(TP+FN) 80.48% and a loss of 0.4263 on train data, and
an accuracy of 84.68% and a loss of 0.3296 on
3.5.3 F1 Score validation data while running it for 15 epochs.

Table 1. Classification report for 15 epochs

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

Figure 9. Confusion Matrix for 15 epochs

Figure 7. Graph of training accuracy vs
validation accuracy for 15 epochs
We performed training with different no. of
epochs. The results for different number of
epochs are shown below:

4.1 Results for 10 epochs

Figure 8. Graph of training loss vs

validation loss for 15 epochs

The initial decrease in both training and

validation loss suggests the model is learning
effectively. However, the validation loss is
plateauing and slightly increasing after the 5th Figure 10. Graph of training accuracy vs
epoch indicates potential overfitting. The validation accuracy for 10 epochs
model might be memorizing the training data's
specifics instead of learning generalizable The green line shows the training accuracy,
features that minimize error on unseen data. which starts at around 72% and quickly
increases to over 80% within the first few
epochs. This means that the model is learning
the training data very well, even memorizing
some of it.

The red line shows the validation accuracy,

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

which starts at around 70% and increases to

around 85% by the 10th epoch. This is lower
than the training accuracy, which suggests that
the model may be over fitting to the training
data and not generalizing well to unseen data.
The gap between the training and validation
accuracy curves suggests the model might be
memorizing the training data instead of
learning generalized features. This could lead
to poor performance on unseen data.

Figure 13. Confusion matrix for 10 epochs

4.2 Results for 20 epoch

Figure 11. Graph of training loss vs

validation loss for 10 epochs

Figure 14. Graph of training accuracy vs

validation accuracy for 20 epochs

Figure 12. Classification report for 10

epochs

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

Figure 17. Confusion Matrix for 20 epochs

Figure 15. Graph of training loss vs
validation loss for 20 epochs 5. CONCLUSION
The utilization of Convolutional Neural
Networks (CNNs) for brain tumor detection in
MRI images exhibits significant potential for
enhancing diagnostic accuracy. The
developed CNN-based model achieved an
impressive 94.1% accuracy in detecting brain
tumors, demonstrating high precision and
recall rates, crucial for distinguishing between
normal brain tissue and tumors.
This project's outcomes indicate CNNs'
viability in aiding radiologists and medical
practitioners in diagnosing brain tumors,
Figure 16. Classification report for 20 potentially enabling earlier and more accurate
epochs diagnoses. However, future research avenues
focus on improving accuracy further and also
This model is experimented with different number detecting the grades of brain tumor by using
of epochs. Although the highest accuracy i.e. 91% segmentation technique such as Malignant,
for 20 epochs, the experimented results for 15 Glioma, Benign. This includes employing
epochs were the best among the experimented more sophisticated architectures such as
epochs. Experimented precision for 15 epochs was resnet50, VGG models, leveraging transfer
highest which predicted the inputs correctly. learning, and using diverse datasets—
incorporating data from other imaging
modalities like CT scans could enhance model
performance.
Moreover, the current model's limitations on
binary classification, overlooking 3D MRI
spatial information, and neglecting real-world
uncertainties highlight areas for improvement.
Future work aims to address these limitations

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

by exploring advanced classification tasks, 6. Evaluation Metrics | 12 Must-Know

developing 3D CNNs, and integrating ML Model Evaluation Metrics. (n.d.).
uncertainty quantification techniques. Available:
https://www.analyticsvidhya.com/blo
6. ACKNOWLEDGMENT g/2019/08/11-important-model-
The authors would like to express the sincere evaluation-error-metrics/
gratitude towards Er.Saban Kumar KC and Er. 7. Havaei, M., Davy, A., Warde-Farley,
Prabin Kumar Jha whose help in completing D., Biard, A., Courville, A., Bengio,
this research is inevitable. Y., Pal, C., Jodoin, P.-M., &
Larochelle, H. (2017). Brain tumor
7. REFERENCES segmentation with Deep Neural
Networks. Medical Image Analysis,
1. Abdusalomov, A. B., Mukhiddinov, M., & 35, 18–31.
Whangbo, T. K. (2023). Brain Tumor 8. Hesaraki, S. (2023, October 18).
Detection Based on Deep Learning Feature Map Medium. Available:
Approaches and Magnetic Resonance https://medium.com/@saba99/feature
Imaging. Cancers, 15(16), 4172. -map-35ba7e6c689e
2. Alzubaidi, L., Zhang, J., Humaidi, A. J., 9. Johny, A., K. N., D. M., & Nallikuzhy,
Al-Dujaili, A., Duan, Y., Al-Shamma, O., D. T. J. (2020). Optimization of CNN
Santamaría, J., Fadhel, M. A., Al-Amidie, Model With Hyper Parameter Tuning
M., & Farhan, L. (2021). Review of deep for Enhancing Sturdiness in
learning: Concepts, CNN architectures, Classification of Histopathological
challenges, applications, future directions. Images (SSRN Scholarly Paper
Journal of Big Data, 8(1), 53. 3735831).
3. Asiri, A. A., Shaf, A., Ali, T., Aamir, M., 10. Lamrani, D., Cherradi, B., el Gannour,
Irfan, M., Alqahtani, S., Mehdar, K. M., O., Bouqentar, M., & Bahatti, L.
Halawani, H. T., Alghamdi, A. H., (2022). Brain Tumor Detection using
Alshamrani, A. F. A., & Alqhtani, S. M. MRI Images and Convolutional
(2023). Brain Tumor Detection and Neural Network. International Journal
Classification Using Fine-Tuned CNN with of Advanced Computer Science and
ResNet50 and U-Net Model: A Study on Applications, 13.
TCGA-LGG and TCIA Dataset for MRI 11. Machine Learning Model
Applications. Life, 13(7). Development and Model Operations:
4. Brain Tumor Detection Using Principles and Practices. (n.d.).
Convolutional Neural Network | IEEE KDnuggets. Available:
Conference Publication | IEEE Xplore. https://www.kdnuggets.com/machine-
(2019). Available: learning-model-development-and-
https://ieeexplore.ieee.org/abstract/docume model-operations-principles-and-
nt/8934561 practices
5. Dertat, A. (2017, November 13). Applied 12. Minz, A., & Mahobiya, C. (2017). MR
Deep Learning - Part 4: Convolutional Image Classification Using Adaboost
Neural Networks. Medium. Available: for Brain Tumor Type. 701–705.
https://towardsdatascience.com/applied- 13. Mustaqeem, A., Javed, A., & Fatima,
deep-learning-part-4-convolutional-neural- T. (2012). An Efficient Brain Tumor
networks-584bc134c1e2

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

Detection Algorithm Using Watershed Dataset Characteristics and Transfer

& Thresholding Based Segmentation. Learning. IEEE Transactions on
International Journal of Image, Medical Imaging, 35(5), 1285–1298.
Graphics and Signal Processing, 4. 19. Supervised Learning: Algorithms,
14. Prathap, P. (2023, July 28). The Secret Examples, and How It Works -
to Understanding CNNs: Convolution, DatabaseTown. (2023, May 26).
Feature Maps, Pooling, and Fully Available:
Connected Layers! Medium. https://databasetown.com/supervised-
Available: learning-algorithms/
https://medium.com/@prajeeshpratha 20. Szegedy, C., Liu, W., Jia, Y.,
p/the-secret-to-understanding-cnns- Sermanet, P., Reed, S., Anguelov, D.,
convolution-feature-maps-pooling- Erhan, D., Vanhoucke, V., &
and-fully-connected-layers- Rabinovich, A. (2015). Going deeper
97055431a847 with convolutions. 2015 IEEE
15. Punia, R. S. (2023, May 6). How to Conference on Computer Vision and
Measure the Performance of Your Pattern Recognition (CVPR), 1–9.
Machine Learning Models: Precision, 21. Team, G. L. (2020, August 17).
Recall, Accuracy, and F1…. Medium. Introduction to Image Pre-processing |
Available: What is Image Pre-processing? Great
https://medium.com/@mycodingmant Learning Blog: Free Resources What
ras/how-to-measure-the-performance- Matters to Shape Your Career!
of-your machine-learning-models- Available:https://www.mygreatlearni
precision-recall-accuracy-and-f1- ng.com/blog/introduction-to-image-
855702df048b pre-processing/
16. Saxena, S. (2021, April 5). 22. Toğaçar, M., Ergen, B., & Cömert, Z.
Introduction to Softmax for Neural (2020). BrainMRNet: Brain tumor
Network. Analytics Vidhya. detection using magnetic resonance
Available: images with a novel convolutional
https://www.analyticsvidhya.com/blo neural network model. Medical
g/2021/04/introduction-to-softmax- Hypotheses, 134, 109531.
for-neural-network/ 23. Umer, A. (2023, August 9).
17. Seth, N. (2021, August 26). Data Understanding Convolutional Neural
Acquisition in Machine Learning—An Networks: A Beginner’s Journey into
Overview. Blogs and updates on Data the Architecture. CodeX. Available:
Science, Business Analytics, and AI https://medium.com/codex/understand
Machine Learning. Available: ing-convolutional-neural-networks-a-
https://www.analytixlabs.co.in/blog/d beginners-journey-into-the-
ata-acquisition/ architecture-aab30dface10
18. Shin, H.-C., Roth, H. R., Gao, M., Lu, 24. Wady, S. H., Yousif, R. Z., & Hasan,
L., Xu, Z., Nogues, I., Yao, J., H. R. (2020). A Novel Intelligent
Mollura, D., & Summers, R. M. System for Brain Tumor Diagnosis
(2016). Deep Convolutional Neural Based on a Composite Neutrosophic-
Networks for Computer-Aided Slantlet Transform Domain for
Detection: CNN Architectures, Statistical Texture Feature Extraction.

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148
Kathford Journal of Engineering and Management,
ISSN: 2661-6106.
© 2023 Kathford Journal of Engineering and Management
https://kathford.edu.np/Journal/
Vol. 3 Issue 1 JUNE 2023, pp.134148

BioMed Research International, 2020,

8125392.
25. What is the AI Life Cycle? - Data
Science Process Alliance. (n.d.).
Available: https://www.datascience-
pm.com/ai-lifecycle/
26. Yamashita, R., Nishio, M., Do, R. K.
G., & Togashi, K. (2018).
Convolutional neural networks: An
overview and application in radiology.
Insights into Imaging, 9(4).

S. Aryal, et al. Kathford Journal of Engineering and Management (KJEM), 2023; 3(1), 134148