0% found this document useful (0 votes)
3 views6 pages

Icosec 226

This paper presents an InceptionV3 infused neural network model for automated brain tumor detection using MRI images, achieving a validation accuracy of 98.50%. The model incorporates advanced data augmentation techniques and a customized training pipeline to enhance feature extraction and classification accuracy. The study demonstrates the model's effectiveness compared to other architectures, emphasizing its potential for reliable medical image analysis.

Uploaded by

productivekunal
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
3 views6 pages

Icosec 226

This paper presents an InceptionV3 infused neural network model for automated brain tumor detection using MRI images, achieving a validation accuracy of 98.50%. The model incorporates advanced data augmentation techniques and a customized training pipeline to enhance feature extraction and classification accuracy. The study demonstrates the model's effectiveness compared to other architectures, emphasizing its potential for reliable medical image analysis.

Uploaded by

productivekunal
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 6

InceptionV3 infused Neural Network for Brain

Tumor Detection
Kunal Bharargava Harshil Nanda *Niha Kamal Basha
School of Computer Science and School of Computer Science and School of Computer Science and
Engineering Engineering Engineering
Vellore Institute of Technology Vellore Institute of Technology Vellore Institute of Technology
Vellore, India. Vellore, India. Vellore, India.
[email protected] [email protected] [email protected]

Abstract— Automated identification and detection of brain [1]. This architecture has been extensively adopted in
tumors using Magnetic Resonance Imaging (MRI) is difficult, medical image analysis because of its strong capability and
time-consuming and laborious because there exists many efficiency. The model's innovative design allows for
overlooked, misread, & deceptive lesions (similar to tumor)
improved feature extraction and classification accuracy,
present in the brain tumor image scans. This paper evaluates
the efficiency of the InceptionV3 infused convolutional neural
making it a reliable choice for medical applications [4]. We
network for brain tumor detection. Utilizing a comprehensive shall be experimenting with different techniques and
dataset Br35H with MRI brain tumor images, initially the methods pertaining to model to maximise our goal and
model preprocessed and augmented the input to enhance achieve high quality results which will later be evaluated
training robustness. The novelty of our approach lies in the and shown via different formats such as graphs and tables.
integration of advanced data augmentation techniques and a This will help us visualise and aid us in our effort to find a
customized training pipeline. The proposed InceptionV3 better and reliable solution. This, however, shall not be
infused neural network model achieved a validation accuracy treated as consolidated, but rather a continuous journey to
of 98.50%, outperforming other models such as EfficientNetB4,
enhance and evaluate the model continuously. The novelty
VGG19, ResNet50, and MobileNetV2 in terms of key metrics.
Our findings demonstrate that the proposed InceptionV3 of our approach lies in the integration of advanced data
infused neural network is highly effective and computationally augmentation techniques with a customized training pipeline
efficient regarding brain tumor detection & classification, tailored specifically for brain tumor classification.
providing one reliable instrument for medical image analysis. Additionally, our method involves fine-tuning a pre-trained
InceptionV3 model with domain-specific modifications to
Keywords— Brain Tumor Detection, InceptionV3, Deep enhance performance. The primary contributions of this
Learning, Convolutional Neural Network, , Data Augmentation, paper include:
Medical Image Analysis, Model Efficiency.

I. INTRODUCTION 1. Implementing advanced data preprocessing and


augmentation techniques to improve the robustness
Brain tumors are among the most aggressive and life- of the training process.
threatening forms of cancer, posing significant challenges to 2. Fine-tuning the pre-trained InceptionV3 infused
healthcare due to their high mortality rate if not detected and neural network model for brain tumor
treated early. Precise and prompt diagnosis assessment of classification.
brain tumor is essential for successful treatment and 3. Comparing the performance of InceptionV3
enhanced patient outcomes. Conventional methods [1] for infused neural network alongside other exceptional
identifying brain tumor, for example magnetic resonance models, including EfficientNetB4, VGG19,
imaging (MRI) and computed tomography (CT) scans, ResNet50, and MobileNetV2.
heavily rely on the expertise of radiologists. The complexity 4. Evaluating the model's efficiency via
and variability of tumor appearances, coupled with the comprehensive metrics like precision, F1-score,
inherent limitations of human interpretation, often lead to recall and accuracy.
delays or inaccuracies in diagnosis [2]. Lately, in latest
years, the advancement of deep learning, more specifically The following below sections of this paper is organized as
on convolutional neural networks (CNNs), etc have brought shown:
a drastic change in analysis of medical image. CNNs have
exhibited significant success in multiple image recognition 1. Literature Survey: Reviews the related work in the
tasks, making them highly suitable for medical applications domain of brain tumor detection by utilizing deep
[3]. These models can automatically learn from various learning models.
sources and extract multiple features from images, enabling 2. Methodology: Details the data collection,
more accurate and consistent analysis compared to preprocessing steps, feature extraction process,
traditional methods [10]. This study harnesses the power of model training, evaluation, and deployment
the InceptionV3 CNN to increase and increase the efficiency procedures.
& accuracy of brain tumor detection and classification. 3. Experimental Results and Analysis: Presents the
results of the experiments, including performance
InceptionV3, developed by Szegedy et al. (2016), utilizes metrics and comparisons with other models.
inception modules to efficiently capture multi-scale features

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


4. Conclusion: Summarizes the discoveries, other domains due to its agility and capacity to capture many
highlighting multiple contributions of the research, complex features in medical images. ResNet50 has been
and suggests possible potential areas for la work. applied in various studies for medical image classification.
For example, its application in the identification of
II. LITERATURE SURVEY pneumonia from X-ray images of chest displayed significant
improvement in accuracy and robustness compared to
MobileNetV2, introduced by Sandler et al., is tailored traditional models [9]. The model's capability to acquire &
for embedded vision & mobile applications [5]. It builds learn multiple complex patterns and its stability during
upon the original MobileNet architecture by incorporating training make it a popular choice for medical image
inverted residuals and linear bottlenecks, improving both analysis.
performance and efficiency. MobileNetV2’s lightweight
architecture makes it highly suitable for applications where InceptionV3, developed by Szegedy et al., utilizes
computational power and memory are limited, such as inception modules to efficiently capture multi-scale features.
portable medical devices for brain tumour detection. This architecture has been extensively implemented in
MobileNetV2 has been effectively used in various medical medical image analysis because of its quick and accurate
imaging applications, including the detection of skin cancer performance and efficiency [10]. The model's innovative
and retinal diseases. Its efficient architecture allows it to design incorporates multiple convolutional filters of multiple
perform well on mobile devices, enabling real-time different sizes enclosed in the same module, empowering it
diagnosis and monitoring in remote and resource-limited to embody a wide range of features at different scales [11].
settings. This multi-scale feature extraction capability allows
InceptionV3 to perform exceptionally well in various
EfficientNet, proposed by Tan et.al, introduces a novel diagnostic tasks, including brain tumor classification [12].
model proportioning method that consistently expands every InceptionV3 has been applied in numerous medical imaging
dimensions of width, resolution & depth by employing a studies. For instance, a study by Sharma et al., demonstrated
compound coefficient. EfficientNetB4, a variant of the the model's effectiveness in detecting brain tumors from
EfficientNet family, achieves state-of-the-art accuracy with MRI images, achieving an accuracy of 97.79% [13]. The
substantially lesser parameters and computational expenses study highlighted the model's ability to handle the high
as contrasted to previous models [6]. The given efficiency variability in tumor shapes and sizes, making it a reliable
makes it an excellent candidate for medical image analysis, tool for medical diagnosis.
including brain tumor detection, where computational
resources may be limited. EfficientNetB4 has been III. METHODOLOGY
successfully used in various medical imaging applications. The methodology for this study focuses on the
For example, its application in the classification of diabetic implementation and evaluation of the InceptionV3 infused
retinopathy images demonstrated superior performance with proposed neural network model for identification of brain
lower computational requirements, making it appropriate for tumor detection and classification. The operational flow of
deployment in many resource-constrained based the proposed work have been shown in Fig. 1.
environments. Its scalability and efficiency are particularly The following steps outline the comprehensive process:
advantageous in medical settings where real-time processing
and accuracy are critical. A. Actual Algorithm
Algorithm on Brain Tumor Detection Using InceptionV3
VGG19, created by Simonyan et.al., is known for its infused neural networks
simplicity and depth, utilizing 19 layers of convolutional Input: Brain tumor image dataset
neural networks [7]. Despite its relatively simple Output: Trained model for brain tumor classification
architecture, VGG19 has shown high efficiency in multiple
image classification functions. The model's deep Step 1: Data Collection
architecture, infused with multiple convolutional layers dataset ← LoadDataset("path/to/dataset")
achieved by multiple max-pooling layers, allows the model
to capture intricate patterns in images, making it suitable for Step 2: Data Preprocessing
applications in medical image analysis, such as brain tumor for each image in dataset do
classification. In a study by Mohsen et al., VGG19 was used image ← Normalize(image)
for the classification of brain tumor images, achieving high image ← Resize(image, 299, 299)
accuracy but facing challenges related to computational image ← Augment(image)
efficiency due to its deep architecture. The study highlighted end for
the need for powerful hardware to leverage the full potential
of VGG19 in medical applications. Step 3: Dataset Splitting
Data is split into 3 sets: train_set, valid_set, test_set ←
ResNet50, developed by Him et al., implemented the idea of SplitDataset(dataset, train_ratio=0.7, valid_ratio=0.2,
residual learning to settle the challenge of dissapearing test_ratio=0.1)
gradients present within multiple deep neural networks [8].
The architecture, with 50 layers, enables the training of Step 4: Model Initialization
much deeper networks by using residual connections, which basic_model ← LoadInceptionV3(weights= “imagenet”,
help to maintain high accuracy. ResNet50 has been widely include_top=False)
implemented widely in various medical image analysis and
new_model ← AddCustomLayers(basic_model, flipping, zooming, shifting, and shearing. Different
num_classes) optimizers like Adam and RMSProp can be utilized to get
different results. By augmenting the data, the model is
Step 5: Compilation exposed to a broader range of image differences, which
compile_model ← Compile(new_model, simulates real-world situations and improves its robustness.
optimizer=“Adam”, loss= “categorical_crossentropy”,
metrics=[“accuracy”]) Feature extraction involves utilizing the pre-trained
InceptionV3 model. This model, which has been re-trained
Step 6: Training on ImageNet, has been loaded and the top classification
history ← TrainModel(compile_model, train_set, layer is removed to fine-tune and refine the model for the
valid_set, epochs=50, batch_size=32, specific tasks of brain tumor classification. Custom layers
callbacks=[EarlyStopping, LearningRateScheduler]) are then integrated to the base model to adjust it to the brain
tumor classification task. This typically includes a Global
Step 7: Evaluation Average Pooling layer which is preceded by dense layers
and a final soft-max layer to output the probability of each
test_results ← EvaluateModel(model, test_set, and every class. The model is then compiled using the
metrics=['precision', 'recall', 'f1-score','accuracy',]) categorical cross-entropy loss, accuracy and RMSProp
optimizer as for evaluation metric. This setup allows the
Step 8: Hyper-parameter Tuning model to effectively learn the intricate features of brain
best_params ← HyperparameterTuning(compile_model, tumor images and make accurate classifications.
train_set, valid_set, param_grid)
The model then is trained on training set defined,
Step 9: Model Testing
alongside the validation set which is used to monitor the
test_metrics ← TestModel(compile_model, test_set)
model’s efficiency. Throughout the training, early stopping
confusion_matrix ← ComputeConfusionMatrix(test_set,
and learning rate scheduling are applied to avoid overfitting
compile_model)
and optimize learning. Early stopping tracks the validation
roc_curve ← ComputeROCCurve(test_set,
loss and halts the training process if there is not
compile_model)
improvement in the loss for a given designated number of
epochs, while learning rate scheduling reduces the learning
Step 10: Deployment
rate whenever a plateau is identified in the performance.
DeployModel(compile_model, "path/to/deployment")
These techniques enhance the training process and
guarantees that the model generalizes adequately to unseen
Step 11: Model Maintenance
data.
while model is in use do
new_data ← CollectNewData()
model ← RetrainModel(compile_model, new_data) Model evaluation is conducted to evaluate the
end while performance of the InceptionV3 infused neural network
end Algorithm model. The model’s performance related to the test set is
evaluated using metrics such as precision, F1-score, recall
and accuracy. These offer a thorough understanding of
The research process begins with data collection,
effectiveness of model in detection and classifying brain
utilizing a diverse and comprehensive dataset of brain tumor
tumor. Additionally, hyperparameter tuning is performed to
images obtained from Kaggle. The dataset, titled "Brain
identify best configuration for the model. Parameters like
Tumor Detection" by Ahmed Hamada, includes a variety of
batch size, number of epochs, learning rate and network
brain tumor types, ensuring a robust model training and
architecture specifics are fine-tuned to achieve the best
evaluation process. It contains MRI images categorized into
possible performance. The model is also evaluated on
four classes: glioma, meningioma, pituitary tumor, and no
obscure data to assess its generalization capability. The
tumor. This variety provides a wide range of examples for
results are analysed using confusion matrices and ROC
the model to learn from, which is crucial for developing a
curves to acquire detailed understanding into the model’s
reliable and accurate classification model.
performance.

Data preprocessing is a critical step focused on


Once we are satisfied and fulfilled with the model’s
preparation the dataset for the proposed InceptionV3 model.
efficiency, the model is implemented for practical use in
This pre-processing involves normalizing the pixel values of
clinical settings. It is necessary to make sure that the model
the given MRI images to a variety of range suitable for the
is integrated with a user-friendly interface for radiologists
InceptionV3 infused neural network model (e.g., [0, 1]),
and healthcare professionals, enabling them to utilize the
which helps fasten up the convergence of neural network.
model effectively for diagnostic purposes.
Additionally, all images given are resized and adjusted to
299x299 pixels, the input size required by the model
architecture. This resizing ensures that the images are Continuous surveillance and monitoring of model’s
compatible with the model's expected input dimensions. To performance [14] is necessary to maintain the accuracy and
further enhance the model’s generalization capability and pertinence. The model is periodically updated with new data
avoid overfitting, different data augmentation techniques are and retrained with updated datasets to adapt to new
implemented. These various techniques incorporate rotation, variations in brain tumor images. This ongoing maintenance
ensures that the model remains effective in real-world
applications and continues to provide accurate diagnostic IV. EXPERIMENTAL RESULTS AND ANALYSIS
support. Here, we demonstrate the experimental results and
analyze the performance-efficiency of the newly-proposed
InceptionV3 infused neural network model on the brain
tumor dataset. The evaluation metrics used include precision,
F1-score, recall and accuracy. The training and validation
process were monitored to measure model's learning &
generalization capabilities.

The model was then trained for 10 epochs, & the


accuracy metrics for both the validation & training phases
were recorded. The training and validation accuracy
improved significantly over the epochs, demonstrating the
model's capability to train from given data effectively. The
final training accuracy is 97.83% and validation accuracy is
98.50%. The graph in Fig.3. shows that the validation
accuracy closely monitors the evaluated training accuracy,
showing that the new model has been generalizing well to
unseen data.

The evaluated loss values for both validation & training


phases were monitored to make sure that model is reducing
the error throughout the training process. The training loss is
0.0731 and validation loss is 0.072. Both validation &
training loss decreased steadily, reflecting the model's
improved performance over the epochs.

The confusion matrix [15] generated shows an in-depth


breakdown analysis of the model's overall performance by
showing the true positive, false positive, true negative and
false negative rates. It reveals that the model correctly
categorised most of the brain tumor images, with a high true
positive and true negative rate, and minimal false positives
and false negatives and are shown in Fig. 2.

Fig. 2. Confusion matrix of proposed model.

Fig. 1. The proposed InceptionV3 infused neural networks workflow


Recall = True Pos. / True Pos. + False Neg. ……..
Eq.3

The models chosen for comparison, including


EfficientNetB4, VGG19, ResNet50, and MobileNetV2, are
widely recognized for their performance in image
classification tasks. Each model symbolizes a distinct
architectural method, allowing a comprehensive evaluation
of the strengths and weaknesses of InceptionV3 against a
diverse set of state-of-the-art models. The study faced
several limitations and challenges, including potential biases
in the dataset, the need for more diverse and larger datasets
Fig. 3. The graph shows the validation & training of the proposed model
over 10 epochs.
to further validate the model, & the computational resources
required for training. These aspects provide a
The new model attains a final training accuracy of comprehensive understanding of the potential areas for
97.83% and a final validation accuracy of 98.50%. The high improvement and future research directions.
accuracy in both validation & training phases show that the
proposed model generalizes quite adequately to obscure data, V. CONCLUSUON
making it highly effective for brain tumor classification..
Table. 1. The table summarizes the key evaluation metrics for the proposed This work validates the effectiveness of the proposed
model InceptionV3 infused neural network for brain tumor
detection and classification. Utilizing advanced data
Table Evaluation Metrics
Head
preprocessing and augmentation techniques, the model
Metric Training Validation
achieved a high validation accuracy of 98.50%,
1 Accuracy 97.83% 98.50% outperforming other models like EfficientNetB4, VGG19,
2 Precision 98.40% 98.40% ResNet50, and MobileNetV2. The combination of a pre-
trained InceptionV3 model with a customized training
3 Recall 98.30% 98.30%
pipeline significantly enhanced its accuracy and robustness.
4 F1-Score 98.35% 98.35% The results highlight InceptionV3's potential as a reliable
Table 2. The performance of the proposed model into perspective, tool for early detection and accurate classification of brain
comparison was done with other deep learning models used in medical tumors, aiding radiologists and healthcare professionals.
image analysis Future work will focus on further optimizing the model,
applying it to other medical images, and integrating it into
Model Comparative Analysis
clinical workflows. Continuous updates and maintenance,
Accuracy Precisio Recall F1-
n Score
along with expanding the dataset to include more diverse
EfficientNetB4 97.25% 97.00% 96.90% 96.95% tumor types, will be essential for maintaining its
VGG19 96.80% 96.60% 96.50% 96.55% effectiveness and relevance. Integration into clinical
ResNet50 98.00% 97.90% 97.80% 97.85% workflows includes developing user-friendly interfaces for
MobileNetV2 97.10% 96.90% 96.80% 96.85% radiologists, ensuring the model's compatibility with
Proposed 98.50% 98.40% 98.30% 98.35% existing medical imaging systems, and conducting pilot
Model studies in clinical settings to validate its utility. The potential
impact includes faster and more accurate diagnosis, which
can significantly improve patient outcomes.
The high precision, F1-score, recall and accuracy in the
above table. 1 suggests that proposed model performs VI. ACKNOWLEDGMENT
exceptionally well in classifying brain tumor images. The The authors extend their heartful gratitude to the VIT
mathematical model [16] for the same has been depicted management for showing support throughout the work.
below in Eq.1, Eq.2, and Eq.3. The above table. 2 shows that
proposed model surpasses other models with regards to REFERENCES
precision, F1-score, recall and accuracy, making it a
superior choice for brain tumor classification tasks. The [1] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens and Z. Wojna,
analysis of the model's performance across different types of "Rethinking the Inception Architecture for Computer
brain tumors reveals how the model performs in detecting Vision," 2016 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016, pp.
gliomas, meningiomas, and pituitary tumors. This analysis 2818-2826, doi: 10.1109/CVPR.2016.308.
provides a deeper understanding of its diagnostic [2] Sharma, M., Sharma, P., Mittal, R., & Gupta, K., “Brain Tumour
capabilities and areas for improvement. Detection Using Machine Learning. Journal of Electronics and
Informatics”, 3(4), 298-308, 2021. DOI:
https://doi.org/10.36548/jei.2021.4.005
Accuracy = True Pos. + True Neg. / True Pos. + False Pos. + [3] Tan, Mingxing and Quoc V. Le. “EfficientNet: Rethinking
True Neg. + False Neg. …………. Eq. Model Scaling for Convolutional Neural
Networks.” ArXiv abs/1905.11946, 2019.
1
[4] Simonyan, Karen and Andrew Zisserman. “Very Deep
Convolutional Networks for Large-Scale Image
Precision = True Pos. / True Pos. + False Pos. …….. Recognition.” CoRR abs/1409.1556, 2014.
Eq.2
[5] He, Kaiming et al. “Deep Residual Learning for Image [12] Emblem, K. E., “Predictive Modeling in Glioma Grading from
Recognition.” IEEE Conference on Computer Vision and MR Perfusion Images Using Support Vector Machines”.
Pattern Recognition,CVPR: 770-778, 2016. Magnetic Resonance in Medicine, 60(4), 945-952, 2018.
[6] Sandler, Mark et al. “MobileNetV2: Inverted Residuals and [13] Rahmani, M., Akbarizadeh, G, “Unsupervised Feature Learning
Linear Bottlenecks.” 2018 IEEE/CVF Conference on Computer Based on Sparse Coding and Spectral Clustering for
Vision and Pattern Recognition: 4510-4520, 2018. Segmentation of Synthetic Aperture Radar Images”. IET
[7] Weiss C. One in Four Dies of Cancer. Questions About the Computer Vision, 9(5), 629-638, 2015.
Epidemiology of Malignant Tumours. Recent Results Cancer [14] Chang, P. D., “Fully Convolutional Deep Residual Neural
Res. 2021;218:15-29. doi: 10.1007/978-3-030-63749-1_2. Networks for Brain Tumor Segmentation”, International
PMID: 34019159. Workshop on Brainlesion: Glioma, Multiple Sclerosis, Stroke
[8] Thrun, S., Pratt, L, “Learning to Learn”. Springer Science & and Traumatic Brain Injuries, 2016.
Business Media, 2012. [15] Ronneberger, O., Fischer, P., Brox, T., “U-Net: Convolutional
[9] Baxter, Jonathan. “Learning to Learn.” Springer US, 1998. Networks for Biomedical Image Segmentation”, International
[10] Beno, M. M., “Threshold Prediction for Segmenting Tumour Conference on Medical Image Computing and Computer-
from Brain MRI Scans. International Journal of Imaging Assisted Intervention,2015.
Systems and Technology, 24(2), 129-137, 2014. [16] Çiçek, Ö., “3D U-Net: Learning Dense Volumetric
[11] El-Dahshan, E. A., “Computer-Aided Diagnosis of Human Segmentation from Sparse Annotation”, International
Brain Tumor through MRI: A Survey and a New Algorithm”. Conference on Medical Image Computing and Computer-
Expert Systems with Applications, 41(11), 5526-5545, 2014. Assisted Intervention,2016.

You might also like