Leakage-Aware Evaluation of a Dual-Backbone Feature-Fusion Ensemble for Multi-Class Brain Tumor MRI Classification (Published)
Deep-learning studies of brain tumor MRI classification often report accuracies close to or above 99%, yet public benchmarks contain duplicate and near-duplicate images that can leak between training and test sets. This study audits the widely used 7,200-image Brain Tumor MRI Dataset and proposes a leakage-aware protocol that removes exact duplicates and groups near-duplicates, using perceptual hashing with pixel correlation, before a group-stratified split. The audit shows that 36.25% of the official test images have a near-duplicate in the official training set. Under the leakage-free protocol, an EfficientNetV2-B0 + ResNet50V2 feature-fusion network and a three-model ensemble were trained with test-time augmentation. The ensemble, selected on validation data alone, achieved 97.80% accuracy (95% CI 96.86–98.64%), macro-F1 of 0.978 and ROC-AUC of 0.998 on 954 unseen images, missing one of 740 tumors; the fusion model alone reached 98.32%. Rigorous de-duplication thus yields credible, still clinically promising, performance estimates.
Keywords: Brain tumor classification, Data Leakage, Ensemble learning, Feature fusion, MRI, transfer learning