ResSGA-Net: A deep learning approach for enhanced brain tumor detection and accurate classification in healthcare imaging systems

Abstract

Accurate and reliable brain tumor classification from magnetic resonance imaging (MRI) is a critical component of computer-aided diagnosis systems, directly impacting clinical decision-making and patient outcomes. This study presents ResSGA-Net, a hybrid deep learning framework that integrates a ResNet50 backbone with dual attention mechanisms (global and gated) and a Swin Transformer to capture both fine-grained local features and long-range contextual dependencies effectively. A fusion strategy is employed to unify convolutional, attention-refined, and transformer-enhanced representations into a robust feature space for multi-class classification. The proposed model is evaluated on two publicly available benchmark datasets, including a four-class and a three-class brain tumor classification task, using stratified cross-validation. Extensive quantitative analysis demonstrates that ResSGA-Net achieves state-of-the-art performance, with accuracies exceeding 98% on Dataset I and strong generalization on Dataset II (accuracy of 93.18% and macro-averaged AUC of 0.989). Comprehensive statistical significance testing confirms that the observed improvements are highly significant and not attributable to random chance. Ablation studies further validate the individual contributions of attention mechanisms and data augmentation strategies, demonstrating that performance gains arise from tumor-specific feature learning rather than artificial data diversity. Qualitative analyses, including confusion matrices, training dynamics, ROC curves, and confidence-based visualizations, confirm stable convergence, robust generalization, and reliable decision confidence across tumor classes. These results indicate that ResSGA-Net provides an accurate, stable, and clinically meaningful solution for automated brain tumor classification, with strong potential for integration into real-world diagnostic imaging workflows.