Hybrid Ensemble Architecture for Brain Tumor Segmentation Using EfficientNetB4-MobileNetV3 with Multi-Path Decoders

Brain tumor segmentation based on multi-modal magnetic resonance imaging is a challenging medical problem due to tumors heterogeneity, irregular boundaries, and inconsistent appearances. For this purpose, we propose a hybrid primal and dual ensemble architecture leveraging EfficientNetB4 and MobileNetV3 through a cross-network novel feature interaction mechanism and an adaptive ensemble learning approach. The proposed method enables segmentation by leveraging recent attention mechanisms, dedicated decoders, and uncertainty estimation techniques. The proposed model was extensively evaluated using the BraTS2019-2021 datasets, achieving an outstanding performance with mean Dice scores of 0.91, 0.87, and 0.83 on whole tumor, tumor core and enhancing tumor regions respectively. The proposed architecture achieves stable performance over a range of tumor types and sizes, with low relative computational cost.