Classification of Brain MRI Images with Vision Transformer: Improving Performance with New Layers and Parameter Optimization

Recent advances in deep learning have revolutionized fields such as robotics, healthcare, and natural language processing. The Vision Transformer (ViT), which operates on self-attention block, has emerged as an alternative to Convolutional Neural Networks. In this study, an improved ViT architecture is proposed for the categorization of brain MRI images as tumorous and non-tumorous. While the standard ViT backbone is utilized for feature encoding, the conventional classification head has been replaced with a multi-stage architecture comprising sequential fully connected layers with 20, 4, and 2 neurons, integrated with Sigmoid and Linear activation functions. This architectural modification, which has been constructed via Greedy Search approach, aims to refine the feature mapping process and enhance the model's sensitivity towards pathological patterns in medical images. To evaluate the performance, a brain MRI dataset has been split into validation, test and training sets, and data augmentation was performed to prevent overfitting. Standard ViT, Swin Transformer, EfficientNet, ResNet and the proposed ViT models have been trained using an ablation technique to optimize network parameters. For the performance analysis, the models have been evaluated using the metrics such as Accuracy, F1-Score, Precision, Recall, AUC values and ROC curves. The results of this work indicate that the proposed ViT’s head structure improves classification success compared to the standard ViT architecture and the other models. Consequently, by redesigning the classification head and optimizing the network, a 13% improvement has been achieved, reaching macro F1-Score of 95.3% for brain image dataset 1 and macro F1-Score of 99.3% for brain image dataset 2.