Enhanced brain tumour detection using fusion of ConvNeXtTiny with stacking generalization

Automated examination of brain Magnetic Resonance Imaging (MRI) remains hampered by class imbalance, high-dimensional feature representations, and limited interpretability, which may compromise the reliability of computer-aided diagnostic systems. Existing research frequently shows good accuracy on publicly available datasets, but it offers limited generalization analysis and inadequate handling of unbalanced data. To address these concerns, this study presents a hybrid pipeline, ConvNeXtTiny-SMOTE-ETC-SM, for binary brain tumor classification (tumor vs. non-tumor) using a publicly available MRI dataset. The framework includes (i) deep feature extraction with ConvNeXtTiny, (ii) data balancing with the Synthetic Minority Oversampling Technique (SMOTE), (iii) feature selection with the Extra Trees Classifier (ETC) to reduce redundancy, and (iv) stacking-based ensemble learning for prediction stability. The fundamental contribution is the systematic incorporation of deep features into conventional ensemble learning and feature selection to overcome class imbalance while preserving interpretability. To ensure robustness, performance was assessed via K-fold cross-validation. The proposed model had an F1-score of 95.9%, a Cohen’s kappa of 91.7%, and a Matthews correlation coefficient (MCC) of 92.0%. Confidence intervals computed with bootstrap resampling revealed performance variability ranging from 89.5 to 100%. While the results are competitive, they are presented without overstating superiority, emphasizing the importance of standardized benchmarking against cutting-edge approaches. From a clinical standpoint, the suggested approach offers a useful tool for improving the reliability of tumor detection in MRI scans, especially in settings with limited expert availability. However, additional validation on multi-institutional datasets and real-world clinical workflows is required prior to deployment.