Hierarchical two-stage ensemble of dilated U-net models for brain tumor segmentation.

Ensemble methods for image segmentation improve performance by combining predictions from multiple models, yielding more accurate and reliable results. This study presents a two-stage hierarchical framework to enhance the accuracy and stability of brain tumor delineation in magnetic resonance imaging data. The proposed approach integrates ensemble strategies at different stages of the processing pipeline. The architecture operates in two stages: first, sub-ensembles resolve internal inconsistencies through simple averaging; second, their outputs are fused into a final prediction using union-based aggregation. The method was evaluated on the Figshare brain tumor dataset and demonstrated progressive performance improvements from individual models to the final hierarchical ensemble. The proposed approach achieved a Dice coefficient of 94.50% and an intersection over union of 89.91%, outperforming existing state-of-the-art methods. The statistical significance of these improvements was confirmed using one-way analysis of variance across three experimental groups, followed by post hoc pairwise testing. The proposed architecture preserves high fidelity in delineating diffuse tumor boundaries and complex morphological structures. By decomposing the ensemble process into two stages, the framework effectively reduces stochastic errors typical of single-model predictions, resulting in a more robust and stable segmentation system that performs reliably even in cases with low contrast and complex tissue interfaces.