Thesis (MSc) - Cyprus International University. Institute of Graduate Studies and Research Electrical and Electronic Engineering

In our current times, with the alarming spread of brain tumors of various types and
causes, the development and improvement of deep models to expedite and simplify
the process of detecting and classifying brain tumors is of utmost importance.
Therefore, in this study, we conduct a comparative analysis of three pretrained deep
learning models VGG19, EfficientNetB3, and ResNet50 with the BTMRI dataset
from Kaggle. The models were implemented and trained using the Google Colab
environment. The evaluation of the models is based on their validation accuracy in
classifying brain tumor images. Initially, the models were trained on this BTMRI
dataset to evaluate their performances which demonstrated varying levels with
validation accuracies of 93.89% for VGG19, 99.09% for EfficientNetB3, and 98%
for ResNet50. Then the performance of these models was enhanced using the Optuna
hyperparameters optimization library to tune three main hyperparameters learning
rate, dropout rate, and number of epochs, due to this the models achieved higher
validation accuracy rates of 95.42% for VGG19, 99.73% for EfficientNetB3, and
99.54% for ResNet50. Overall This study highlights the effectiveness of deep
learning models in medical image classification and the impact of hyperparameter
optimization in improving model performance. These findings could guide future
research and practical applications in the field of medical diagnostics, contributing to
the detection of brain tumors more accurately and efficiently.