Data Efficient Vision Transformer Based Model for Brain Tumor Classification in MRI
2821-0263
Abstract
Accurate detection of brain tumors at early stages is crucial for clinical diagnosis and treatment planning. Magnetic Resonance Imaging (MRI) is the most popular method due to its soft-tissue contrast and spatial resolution. However, manual MRI interpretation is time-consuming and can be inconsistent. Traditional computer-aided diagnosis systems that use convolutional neural networks (CNNs) have some limitations. Vision Transformer (ViT) models have emerged as a viable alternative, leveraging self-attention mechanisms for long-range spatial dependencies. This study introduces an automated brain tumor detection framework based on a pre-trained ViT model to classify MRI images as tumor or non-tumor. The methodology includes a pre-processing framework for resizing, intensity normalization, and data augmentation to enhance model performance and diversity. The ViT model employs a transfer learning approach with ImageNet-pretrained weights, replacing its original classification head with a binary-focused layer. A class-weighted cross-entropy loss function addressed dataset imbalance with model training optimized using carefully selected parameters. Performance metrics such as accuracy, sensitivity, specificity, and receiver operating characteristic area under the curve (ROC-AUC) evaluate the model on a public brain MRI dataset. Results indicate that the framework achieves a remarkable accuracy of 95.12% with a test accuracy of 90.20% and a high ROC-AUC score of 0.9274, demonstrating its discriminative power. The framework achieves 100% sensitivity in tumor classification and eliminates false negatives, a vital feature for clinical applications. The ViT model proves superior by effectively capturing global contextual information and addressing CNN limitations in modeling long-range relationships. This system has significant potential in hospital and diagnostic settings, improving diagnostic accuracy and patient outcomes.