A Survey of Methods and Architectures for Brain MRI Image Classification for Cancer Detection Using Transformer and Group Parallel Axial Attention with Quantum Self-Attention

Abstract

Brain tumor classification using Magnetic Resonance Imaging (MRI) has emerged as a crucial application of artificial intelligence in healthcare, enabling early and accurate diagnosis. Traditional methods based on Convolutional Neural Networks (CNNs) have demonstrated strong performance in extracting local spatial features; however, their limited capability to capture global dependencies restricts effectiveness in complex classification tasks. Recent advancements have introduced Transformer-based architectures, axial attention mechanisms, and hybrid deep learning models that significantly improve classification accuracy and computational efficiency. Transformers utilize self-attention to model long-range relationships within MRI data, enhancing feature representation, while hybrid CNN-Transformer models combine local and global feature extraction for superior performance. Axial attention further optimizes computation by decomposing attention operations, making it suitable for high-resolution images. Additionally, quantum self-attention presents an emerging paradigm by incorporating quantum-inspired principles to enhance learning efficiency. Comparative analyses indicate that attention-based and hybrid models achieve very high classification accuracy. Despite these advancements, challenges such as limited data availability, high computational cost, and lack of interpretability persist, highlighting the need for lightweight, explainable, and clinically deployable solutions.