Diagnosis of brain tumor using transfer learning and deep learning methods

Brain tumors are serious health problems that require early diagnosis. In recent years, artificial intelligence-based methods have been widely used to detect tumors from medical imaging data. However, working with datasets that contain a limited number of samples makes accurate and reliable diagnosis difficult. In this thesis study, deep learning models based on transfer learning were developed using VGG16, InceptionV3, ResNet50, EfficientNetB0, and MobileNetV2 architectures to detect the presence of brain tumors from magnetic resonance (MR) images. In all models, customized layers were added on top of the pre-trained structures; to achieve stronger generalization performance with limited data, data augmentation and fine-tuning techniques were applied. During the training and validation stages, the open-source "Brain Tumor Dataset" accessed from the Kaggle platform was used; in the testing stage, in addition to this dataset, samples from the "Brain Tumor Classification through MRI" dataset were also included. According to the evaluation results, the models that achieved the highest performance were VGG16 and InceptionV3. The VGG16-based model demonstrated high detection accuracy and generalization performance with 99.99% training accuracy, 99.00% validation accuracy, and a 99.75 F1-score. Similarly, the InceptionV3-based model achieved strong results with 99.87% training accuracy, 99.50% validation accuracy, and a 99.75 F1-score. In both models, all negative samples were correctly predicted, and only one false negative prediction was made in the positive class; this indicates that the models provided balanced and reliable performance in terms of sensitivity and specificity. With this study, the performances of models based on different transfer learning architectures were compared, and two distinct models that achieved successful results in the automatic, accurate, and early detection of brain tumors were introduced to the literature.