An advanced deep learning approach for medical image analytics in brain tumor diagnosis

Brain tumors remain difficult to detect at an early stage, and reliable detection on magnetic resonance imaging (MRI) scans is hindered by dataset variability, subtle inter-class similarities, and the limited generalization of existing deep learning models. Despite recent advances in deep learning–based brain tumor analysis, many studies rely on limited datasets and simplified evaluation settings, offering restricted insight into how different architectural paradigms exploit spatial, sequential, and global contextual information under clinically realistic conditions. To address this gap, this study introduces a rigorous multi-paradigm analytical framework that systematically contrasts convolutional, recurrent, and Transformer-based learning strategies under identical preprocessing, training, and evaluation conditions to examine how spatial, sequential, and global feature representations influence brain tumor classification performance. Specifically, Convolutional Neural Networks (CNNs), Long Short-Term Memory networks (LSTMs), Vision Transformers (ViTs), and hybrid architectures are evaluated within a unified experimental setting. A two-stage diagnostic pipeline is formulated in which tumor presence detection precedes subtype classification (glioma, meningioma, pituitary), enabling consistent and fair comparison across model families while reflecting real-world radiological diagnostic workflows. Within this framework, the effectiveness of attention-augmented hybrid architectures is evaluated by integrating temporal modeling and attention mechanisms into convolutional backbones, yielding clear gains in feature selectivity and diagnostically relevant region localization. Classical CNN models achieve accuracies ranging from 97.62% for binary classification to 90.51% for multi-class tasks, while CNN–LSTM hybrids improve performance to 98.88% and 96.89%, respectively. The incorporation of attention mechanisms further increases accuracy to 99.85% for binary classification and 99.21% for multi-class classification. Transformer-based models deliver the strongest overall performance, with Vision Transformer variants achieving 99.92% accuracy for binary detection and 96.49%-96.91% for multi-class classification, highlighting the benefits of global contextual modeling for MRI-based tumor analysis. • Develop an attention-based model for binary and multi-class brain tumor classification. • Analyze MRI brain images to identify different tumor categories in healthcare applications. • Integrate convolutional, recurrent, and attention layers for enhanced medical image analytics. • Achieve high classification accuracy supporting reliable diagnostic decision-making. • Demonstrate how analytics improves healthcare outcomes in brain cancer detection.