Automatic classification of brain tumors in MR images using convolutional neural network and vision transformer architectures

The brain is the most important organ for humans and is the main organ of the central nervous system. The brain manages basic human activities such as thinking, movement and senses. Brain tumors, characterized by the uncontrolled growth of cells in brain tissue, pose a serious threat to both the functioning of the central nervous system and general health. Accurate classification of brain tumor types is crucial for determining appropriate treatment methods and improving patient outcomes. Recent advances in artificial intelligence (AI) and deep learning (DL) have demonstrated significant potential for automatic classification of brain tumors using medical imaging data. This study aims to use AI and DL approaches to develop an automated system for classifying brain tumor types from magnetic resonance (MR) images. In this research, convolutional neural networks (CNN) and vision transducers (ViT) were used to classify brain tumor types using the features of a total of 32490 images from three different publicly available datasets. To ensure robust model evaluation, the datasets were divided into 80% training, 10% validation and 10% test sets, respectively. The methodology uses three different ViT models (ViT-B/16, ViT-L/16 and ViT- H/14) pre-trained on ImageNet-21k dataset and one CNN model (BiT-50) pre-trained on ImageNet-21k dataset and transfer learning for feature extraction. Performance measurements of the deep learning models when accurately classifying different types of brain tumors showed that the models achieved high classification accuracy on all datasets. In particular, the ViT-B/16 model achieved 98.8% accuracy in dataset I, 97.08% in dataset II and 96% in dataset III. The ViT-L/16 model achieved 99.5% accuracy in dataset I, 97.55% in dataset II and 97.54% in dataset III. ViT- H/14 model achieved 97.3% accuracy in dataset I, 95.89% accuracy in dataset II and 93.23% accuracy in dataset III. The BiT-50 model achieved 99.9% accuracy in dataset I, 98.64% accuracy in dataset II and 99.68% accuracy in dataset III. As a result, this research emphasizes the successful application of CNN and ViT models for automatic classification of brain tumor types from brain MRI images. Future work will focus on improving data augmentation techniques, incorporating more different preprocessing methods, and exploring new deep learning models to further improve classification accuracy.