Classification of brain tumors from magnetic resonance images with deep learning methods

In this thesis, a deep learning–based diagnostic support system for the automatic classification of brain tumors from magnetic resonance imaging (MRI) images was developed, and the performances of different convolutional neural network (CNN) architectures were comparatively evaluated. The study employed the Brain Tumor MRI Dataset published on the Kaggle platform, which contains a total of 7,023 images belonging to four classes: glioma, meningioma, pituitary, and no tumor. Within the scope of the research, the MobileNetV2, EfficientNetB0, VGG16, ResNet50, and DenseNet121 models included in the Keras Applications library were trained using the transfer learning method. In addition, an original CNN architecture was designed as part of the thesis, and a hybrid architecture was developed by combining the feature extraction capabilities of MobileNetV2 with those of this original CNN model. All models were compared in a fair manner by employing the same preprocessing steps, data augmentation techniques, and training parameters. According to the standard training results, the highest accuracy was obtained by the original CNN model with 90.31%, while the hybrid model ranked second with an accuracy of 85.51%. The ResNet50, MobileNetV2, and DenseNet121 models produced accuracy values of 83.79%, 80.40%, and 79.63%, respectively. The VGG16 model achieved 73% accuracy, whereas the EfficientNetB0 model exhibited the lowest performance with only 30.89% accuracy. To evaluate model generalizability, five-fold cross-validation was applied. Based on these results, the hybrid model became the most successful structure among all models with an average accuracy of 93.15% and an F1-score of 93.08%. The original CNN model ranked second with an average accuracy of 92.72%. The ResNet50, MobileNetV2, DenseNet121, and VGG16 models achieved average accuracy values of 91.20%, 89.12%, 87.74%, and 82.28%, respectively. The EfficientNetB0 model achieved only 27.45% accuracy in cross-validation as well. This indicates significant convergence difficulties of the corresponding architecture in capturing the specific textural characteristics of the brain tumor dataset. The obtained findings reveal that the original CNN and especially the hybrid architecture offer high accuracy, low variance, and strong generalization capability in MRI-based brain tumor classification problems. This study provides a reliable basis for the development of clinical decision support systems.