An adaptive hierarchical class-aware deep ensemble strategy for robust brain tumor classification
Automated magnetic resonance imaging (MRI) classification can support differentiation of brain tumor categories, but conventional ensembles often depend on opaque probability fusion. This study proposes a validation-driven adaptive hierarchical class-aware ensemble that treats VGG19 and Darknet53 as complementary experts. A balanced working set of 7,200 images representing glioma, meningioma, no tumor, and pituitary classes was stratified into training (70%), validation (15%), and independent test (15%) partitions. Training-only enhancement and best-validation-checkpoint retention were used. Validation data alone produced a frozen expert map: Darknet53 was assigned to glioma, meningioma, and no-tumor cases, whereas VGG19 was assigned to pituitary cases. At test time, the router accepted model consensus, consulted the expert map during disagreement, and used confidence only for residual conflicts. On 1,080 unseen test images, the proposed strategy correctly classified 1,060 cases (98.15% accuracy; 98.16% macro precision; 98.15% macro recall; 98.14% macro F1-score). Darknet53 and VGG19 achieved 97.87% and 97.69% accuracy, while weighted soft voting, and stacking each achieved 98.06%. The proposed strategy therefore achieved the strongest benchmark while preserving a transparent, auditable, and leakage-free decision pathway.