Research Paper Minor Project Final
Research Paper Minor Project Final
Research Paper Minor Project Final
MRI Images
Saday Samnotra Avishkaar Chaturvedi Tanmay Sehgal Chayan Manocha
Computer Science Computer Science Computer Science Computer Science
Bharati Vidyapeeth’s College of Bharati Vidyapeeth’s College Bharati Vidyapeeth’s College Bharati Vidyapeeth’s College
Engineering of Engineering of Engineering of Engineering
Delhi, India Delhi, India Delhi, India Delhi, India
[email protected] [email protected] [email protected] [email protected]
Amrita Ticku
Asst. Proffessor,Computer Science
Bharati Vidyapeeth’s College of
Engineering
Delhi, India
[email protected]
Abstract— The study looks at the use of transfer learning, cutting-edge machine learning algorithms and modern
namely EfficientNet, to classify various forms of brain tumours imaging technology.
using a curated set of T1, contrast-enhanced T1, and T2
magnetic resonance images (MRI). With tumours ranging from Our dataset contains a comprehensive collection of T1,
astrocytoma to tuberculoma, the study intends to break down contrast enhanced T1, and T2 magnetic resonance images that
traditional diagnostic hurdles using artificial intelligence, have been rigorously categorised according to numerous brain
providing insights into subtle tumour subgroups. The study tumour forms, such as astrocytoma, carcinoma, ependymoma,
aims to improve diagnostic precision, prognostic insight, and and others. These pictures, collected without the patient's
treatment efficacy in the complex field of neuro-oncology by identify, give vital insights for neuroimaging research and
utilising cutting-edge neuroimaging and computational analyses.
approaches.
EfficientNet is a revolutionary approach to convolutional
Keywords—Efficient Net, transfer learning, neuro-oncology, neural network (CNN) architectures that emphasises efficient
Astrocytoma, tubercloma. model scaling and enhanced performance. By combining
EfficientNet V2 B0 and ViT-B16 into our classification
framework, we are able to accelerate model training while also
improving classification accuracy. Ensembling strategies
I. INTRODUCTION
improve the resilience of our models, allowing for complete
The human brain, one of the most important organs in the analysis and interpretation of brain imaging data.
body, can encounter considerable obstacles as a result of
abnormal brain cell proliferation, which can cause serious Our approach differs from prior efforts by pioneering the
harm. The frequency of brain tumours is increasing, needing integration of cutting-edge transfer learning models, such as
new diagnostic and therapeutic techniques. A brain tumour is EfficientNet, with classic image and data mining techniques.
characterised by an abnormal aggregation of cells within the This novel method has the potential to revolutionise
skull and presents a complicated medical challenge. These neuroimaging by allowing for more accurate and efficient
tumours are classified as normal, malignant, or benign, with diagnosis of brain tumours. We aim to create substantial
primary or secondary causes. Primary tumours form within advances in neuro-oncology medical research and patient
the brain from a variety of brain cells and tissues, whereas treatment by merging cutting-edge technology with advanced
secondary tumours spread throughout the body. Early machine learning techniques.Press the enter key to begin a
diagnosis of brain tumours is crucial for successful treatment new paragraph. The proper spacing and indent are
and avoiding consequences. automatically applied.
II. LITERATURE REVIEW
In response to the growing need for reliable brain tumour
identification and classification, we offer an automated The combination of several computational algorithms and
classification strategy based on image and data mining medical imaging modalities has resulted in substantial
approaches. Our goal is to investigate the efficacy of transfer advances in brain cancer detection and classification
learning models, notably EfficientNet V2 B0 and ViT-B16, research. Numerous research have contributed to improving
paired with ensembling approaches for identifying various the accuracy and efficiency of brain tumour detection
forms of brain tumours. This project intends to better the systems, using approaches ranging from standard machine
diagnosis process and improve patient outcomes by utilising learning algorithms to cutting-edge deep learning
architectures. with medical imaging holds substantial potential for
transforming the diagnosis and treatment of brain cancer.
Banerjee et al. [1] emphasised the importance of noninvasive
techniques such as MRI in brain cancer detection and III. PROPOSED METHODOLOGY
proposed an adaptive neuro-fuzzy classifier based on
linguistic hedges (ANFC-LH) for simultaneous feature
selection and tumour grade prediction, with a significantly
higher testing accuracy of 85.83%.
A) IMAGE PRE-PROCESSING
The preprocessing phase of brain tumor images stands as a
pivotal component within our methodology, strategically
engineered to refine data representation and expedite model
training and evaluation processes. Through harnessing the
versatile capabilities of TensorFlow's tf.data API, we
proficiently load and manage the dataset, fostering seamless
handling of extensive image data. Employing batching Figure 2: Image of MRI after augmentation and preprocessing
techniques enables concurrent processing of multiple images,
thereby harnessing parallel processing capabilities to bolster
Within our methodology, data augmentation techniques, We endeavor to assess the efficacy of our scaling
notably random flipping and zooming, are strategically methodology by juxtaposing it against established
applied to augment images. Random flipping entails both Convolutional Neural Networks (ConvNets). However, to
horizontal and vertical mirroring of images with a defined underscore the potency of our scaling methodology, we have
probability, fostering the model's acquisition of invariant devised a novel baseline tailored to mobile-size constraints,
features and orientations, thereby enhancing its adaptability christened EfficientNet V2.
to images with diverse orientations or viewpoints.
EfficientNet V2 epitomizes a highly efficient convolutional
Likewise, random zooming introduces fluctuations in the neural network (CNN) architecture lauded for its exceptional
scale and perspective of images through zooming in or out performance and scalability across an array of computer
with a designated factor. This augmentation modality vision tasks. Pre-trained on expansive image datasets, the
simulates fluctuations in image resolution and scale, architecture engenders a rich repository of hierarchical visual
equipping the model with improved capabilities to address representations, ripe for fine-tuning to our specific objective
images of disparate sizes and resolutions encountered within of brain tumor classification.
real-world contexts.
Our overarching aim converges on maximizing model
Through the integration of data augmentation into the accuracy within defined resource constraints, thereby
preprocessing pipeline, our objective is to immerse the model constituting an optimization conundrum warranting
in a multifaceted array of image variations and scenarios, meticulous scrutiny and strategic resolution.:
mirroring the complexity inherent in real-world data. This
strategic maneuver serves as a bulwark against overfitting by
infusing noise and variability into the training data, thereby
thwarting the model's inclination towards memorizing
Figure 3: Distribution of total images inro their respective criteria. Table 1Efficient Net Baseline Data
Furthermore, the F1-score, representing the harmonic mean V) CONCLUSION AND FUTURE SCOPE
of precision and recall, offers a balanced assessment of the In conclusion, the application of deep learning
model's performance. With an F1-score of 91.63%, our model methodologies, particularly convolutional neural networks
showcases a robust equilibrium between precision and recall, (CNNs), has exhibited considerable promise within the realm
underscoring its capacity to sustain elevated levels of both of brain tumor classification utilizing MRI images. Through
metrics concurrently. the implementation of an efficient net architecture and
diligent training methodologies, we have showcased the
In tandem with individual metrics, the macro and weighted capacity of deep learning models to effectively discern brain
averages furnish a comprehensive overview of the model's tumor subtypes, yielding commendable precision and recall
performance across all classes. The macro average, scores. The holistic evaluation metrics, encompassing
computing the average of metrics for each class sans precision, recall, F1-score, and accuracy, underscore the
consideration of class imbalance, yields an average F1-score efficacy of our model in delineating between distinct tumor
of 84%, while the weighted average, accounting for class classes, thereby furnishing invaluable insights for clinical
imbalance, yields an average F1-score of 92%. diagnosis and treatment strategizing.
Class Precision Recall F1- Support
Score
Astrocitoma T1 1.00 0.92 0.96 13 The successful integration of data preprocessing techniques,
Astrocitoma 0.92 1.00 0.96 11 inclusive of image augmentation and input data pipeline
T1C+ construction, has emerged as a pivotal factor in fortifying the
Astrocitoma T2 1.00 1.00 1.00 9 resilience and generalization prowess of the model. By
Ependimoma 1.00 1.00 1.00 4 incorporating these preprocessing paradigms, we have
T1C+ adeptly mitigated challenges associated with data scarcity,
Ependimoma T2 1.00 1.00 1.00 4 class imbalance, and overfitting, thereby fortifying the
Ganglioglioma 1.00 1.00 1.00 2 model's adaptability to variances in input data and
T1 augmenting its overall efficacy.
Ganglioglioma 0.50 1.00 0.67 1
T1C+ Looking forward to a plethora of avenues for future research
Ganglioglioma 1.00 1.00 1.00 3
and refinement within the domain of brain tumor
T2
classification using deep learning beckon. Primarily, the
Germinoma T2 1.00 1.00 1.00 4
Meduloblastoma 1.00 1.00 1.00 2 exploration of advanced neural network architectures and
T2 ensemble methodologies could further amplify the model's
Meningioma T1 1.00 1.00 1.00 28 predictive aptitude and resilience to noise and variations in
Meningioma 1.00 0.96 0.98 56 MRI images. Furthermore, the incorporation of multimodal
T1C+ imaging data, encompassing functional MRI (fMRI) and
diffusion tensor imaging (DTI), harbors potential for [https://www.semanticscholar.org/paper/Brain-tumor-detection-and-
classification-using-a-Amin-
capturing complementary insights into tumor attributes and Sharif/f630ee33272b8505668b70975c3c9df015b21105].
spatial interrelationships, thereby fostering more nuanced and
precise tumor classification frameworks.
[7] Y. Bhanothu, A. Kamalakannan and G. Rajamanickam, "Detection and
Classification of Brain Tumor in MRI Images using Deep
Additionally, the deployment of transfer learning strategies, Convolutional Network," 2020 6th International Conference on
whereby pre-trained models are fine-tuned on bespoke brain Advanced Computing and Communication Systems (ICACCS),
tumor datasets, stands poised to expedite model development Coimbatore, India, 2020, pp. 248-252, doi:
10.1109/ICACCS48705.2020.9074375. keywords:
and facilitate knowledge transfer across diverse medical {Tumors;Classification algorithms;Proposals;Image
imaging endeavors. Moreover, the integration of segmentation;Training;Magnetic resonance imaging;Testing;Brain
interpretability techniques, such as attention mechanisms and tumor detection;glioma;meningioma;pituitary;Faster R-CNN;VGG-
saliency map, has the potential to furnish clinicians with 16;mean average precision;MRI},
invaluable insights into the decision-making mechanisms of
the deep learning model, fostering trust and confidence in its [8] D. M. Joshi, N. K. Rana and V. M. Misra, "Classification of Brain
Cancer using Artificial Neural Network," 2010 2nd International
diagnostic acumen. Conference on Electronic Computer Technology, Kuala Lumpur,
Malaysia, 2010, pp. 112-116, doi: 10.1109/ICECTECH.2010.5479975.
In summation, the convergence of deep learning keywords: {Artificial neural networks;Neoplasms;Magnetic resonance
methodologies and medical imaging evinces substantial imaging;System testing;Cancer detection;Feature extraction;Computer
networks;Lesions;Image processing;Histograms;Brain
potential for revolutionizing the landscape of brain tumor cancer;MRI;Co-occurrence Matrix;ANN},
diagnosis and treatment. Through the adept utilization of
cutting-edge techniques and the cultivation of [9] H. T. Zaw, N. Maneerat and K. Y. Win, "Brain tumor detection based
interdisciplinary collaborations, we stand poised to usher in on Naïve Bayes Classification," 2019 5th International Conference on
an era marked by more precise, efficient, and accessible Engineering, Applied Sciences and Technology (ICEAST), Luang
healthcare solutions for patients worldwide.. Prabang, Laos, 2019, pp. 1-4, doi: 10.1109/ICEAST.2019.8802562.
keywords: {Tumors;Magnetic resonance imaging;Image
segmentation;Feature extraction;Cancer;Classification
algorithms;Databases;Glioblastoma Multiforme (GBM);Magnetic
VI) REFERENCES resonance imaging (MRI);Naïve Bayes classification;maximum
[1] S. Banerjee, S. Mitra and B. U. Shankar, "Synergetic neuro-fuzzy entropy threshold;statistical features extraction},
feature selection and classification of brain tumors," 2017 IEEE
International Conference on Fuzzy Systems (FUZZ-IEEE), Naples, [10] P. Thirumurugan and P. Shanthakumar, "Brain tumor detection and
Italy, 2017, pp. 1-6, doi: 10.1109/FUZZ-IEEE.2017.8015514. diagnosis using ANFIS classifier," International Journal of Imaging
keywords: {Feature extraction;Tumors;Pragmatics;Magnetic Systems and Technology, vol. 26, 2016. [Online]. Available:
resonance imaging;Cancer;Shape}, https://api.semanticscholar.org/CorpusID:40434958
[11] S. Chauhan, A. More, R. Uikey, P. Malviya and A. Moghe, "Brain
[2] M. Gurbină, M. Lascu and D. Lascu, "Tumor Detection and tumor detection and classification in MRI images using image and data
Classification of MRI Brain Image using Different Wavelet mining," 2017 International Conference on Recent Innovations in
Transforms and Support Vector Machines," 2019 42nd International Signal processing and Embedded Systems (RISE), Bhopal, India,
Conference on Telecommunications and Signal Processing (TSP), 2017, pp. 223-231, doi: 10.1109/RISE.2017.8378158. keywords:
Budapest, Hungary, 2019, pp. 505-508, doi: {Feature extraction;Tumors;Image segmentation;Image color
10.1109/TSP.2019.8769040. keywords: {Tumors;Support vector analysis;Image edge detection;Data mining;Clustering
machines;Magnetic resonance imaging;Noise reduction;Discrete algorithms;MRI image;L∗a∗b colour space;K-means
wavelet transforms;brain;classification;denoising;detection;support clustering;segmentation;Classification},
vector machines;tumor;wavelet transforms},
[3] M A Ansari, Rajeev Agrawal, and Rajat Mehrotra, “Detection and [12] A. Rehman, M. A. Khan, T. Saba, Z. Mehmood, U. Tariq, and N.
classification of brain tumor in MRI images using wavelet transform Ayesha, “Microscopic brain tumor detection and classification using
and support vector machine,” 2002. 3D CNN and feature selection architecture,” Microscopy Research and
Technique, vol. 84, no. 1, pp. 133–149, Sep. 2020, doi:
[4] A. Hebli and S. Gupta, "Brain tumor prediction and classification using https://doi.org/10.1002/jemt.23597.
support vector machine," 2017 International Conference on Advances
in Computing, Communication and Control (ICAC3), Mumbai, India, [13] [13]A. Hu and N. Razmjooy, "Brain tumor diagnosis based on
2017, pp. 1-6, doi: 10.1109/ICAC3.2017.8318767. keywords: metaheuristics and deep learning," International Journal of Imaging
{Tumors;Feature extraction;Support vector machines;Discrete wavelet Systems and Technology, vol. 31, pp. 657-669, 2020. [Online].
transforms;Training;Magnetic resonance imaging;Principal Available: https://api.semanticscholar.org/CorpusID:224879353
component analysis;Brain Tumor;MRI (magnetic resonance
imaging);DWT (Discrete wavelet transform);PCA (Principal
[14] M. Thachayani and S. Kurian, “AI Based Classification Framework
component analysis);SVM (Support vector machine)},
For Cancer Detection Using Brain MRI Images,” 2021 International
Conference on System, Computation, Automation and Networking
[5] J. Amin, M. Sharif, A. Haldorai, M. Yasmin, and R. S. Nayak, "Brain (ICSCAN), Jul. 2021, doi:
tumor detection and classification using machine learning: a https://doi.org/10.1109/icscan53069.2021.9526456.
comprehensive survey," Complex & Intelligent Systems, vol. 8, pp.
3161-3183, 2021. [15] Pritanjli and Amit Doegar, “Brain Tumor Detection and Classification
Using Machine Learning,” Algorithms for intelligent systems, pp. 227–
[6] T. Sadad, A. Rehman, A. Munir, Tanzila, U. Tariq, Noor, and R. A. 234, Jan. 2020, doi: https://doi.org/10.1007/978-981-15-0426-6_24.
Rehman, "Brain tumor detection and multi-classification using
advanced deep learning techniques,". Available: