Multi-Class Classification of Brain Cancer Based on MRI Images Using MobileViT Architecture

Abstract

Brain cancer is a disease with serious impacts on neurological functions and patients’ quality of life. Early detection via MRI is crucial, yet manual image interpretation by radiologists is subjective and prone to errors. While deep learning approaches offer automation solutions, conventional models are often constrained by high computational burdens, making them difficult to implement on resource-constrained devices. Addressing these challenges, this study develops an efficient multi-class classification system using the Mobile Vision Transformer (MobileViT) architecture to detect four conditions: glioma, meningioma, pituitary, and normal within MRI datasets. The research methodology encompasses three systematic experimental scenarios covering hyperparameter optimization via grid search, comparative analysis of fine-tuning strategies, and evaluation of the impact of geometric data augmentation on model generalization. Performance evaluation is conducted based on accuracy, precision, recall, and F1-score metrics. The results indicate that the configuration of the AdamW optimizer, a learning rate of 0.001, and a dropout of 0.2 provides optimal training stability. Significant findings reveal that the fine-tuning strategy is crucial for medical feature adaptation, achieving an accuracy surge to 97.40%. Furthermore, the application of data augmentation successfully reduced prediction bias and strengthened the model’s robustness against input variations, increasing the accuracy to 98.70%. These results demonstrate that MobileViT is capable of delivering superior diagnostic performance while maintaining computational efficiency.