Efficient and explainable MRI brain tumor classification via Adaptive GAN-based data augmentation

Accurate brain tumor diagnosis from MRI is critical, but state-of-the-art deep learning (DL) models are computationally expensive and require large datasets, limiting clinical deployment. Classical machine learning (ML) is more efficient but struggles with imbalanced data and a persistent performance gap. This study proposes an XAI-driven Adaptive GAN (AGAN) framework to bridge this gap. The AGAN integrates interpretability scores (e.g., SHAP, Grad-CAM) into its training objective to iteratively refine synthetic sample generation, focusing on improving minority tumor class representation. Evaluated on four diverse public MRI datasets, the AGAN framework demonstrated a 19.7% improvement in minority-class F1-score over the baseline. AGAN-augmented classical ML classifiers, particularly Support Vector Machine (SVM), achieved accuracies up to 0.902, narrowing the performance gap to state-of-the-art DL models (e.g., ResNet50) to within 2.8–5.7%. Critically, the ML pipeline delivered up to 7000 × faster training and required up to 240 × less memory. These findings validate the XAI-guided, AGAN-augmented ML pipeline as a scalable, resource-efficient, and interpretable alternative for brain tumor diagnosis, highly suitable for deployment in resource-constrained clinical environments. • XAI-guided Adaptive GAN using interpretability scores to iteratively refine synthetic sample generation for minority tumor classes. • GAN-augmented classical ML models achieve up to 90% accuracy, closing the gap to state-of-the-art DL models to within 3%–6% while being up to 7000 × faster and 240 × more memory-efficient. • Multi-dataset validation shows 19.7% F1-score improvement in minority-class detection over standard augmentation. • Clinically interpretable pipeline enabling transparent brain tumor diagnosis in resource-constrained environments.