An Explainable Hybrid Deep Learning Framework for Brain Tumor Segmentation and Classification Using U-Net and EfficientNet-B7

Abstract

Accurate brain tumor localization and tumor/non-tumor classification from multimodal magnetic resonance imaging (MRI) remain challenging because standalone segmentation and classification models provide complementary but limited information for comprehensive tumor analysis. To address this limitation, a hybrid deep learning framework integrating U-Net and EfficientNet-B7 is proposed. The framework employs U-Net to generate pixel-level tumor segmentation masks from multimodal MRI sequences (T1, T1ce, T2, and FLAIR), followed by tumor region-of-interest (ROI) extraction for EfficientNet-B7-based tumor/non-tumor classification. The segmentation confidence and classification probability are subsequently combined using a weighted probabilistic fusion strategy, with the fusion weight experimentally determined during model development. Experiments were conducted on the BraTS 2023 dataset using subject-wise partitioning into training, validation, and testing subsets at a ratio of 70%:15%:15%, and the experiments were repeated five times using different random initialization seeds. The proposed framework achieved a Dice Similarity Coefficient (DSC) of 0.91 and a Jaccard index of 0.86 for segmentation, while achieving 96.8% accuracy, 95.8% sensitivity, 97.5% specificity, 95.5% F1-score, and 0.97 AUC for classification. Comparative evaluation demonstrated higher overall performance than the evaluated baseline architectures, including U-Net, V-Net, ResNet50, and EfficientNet-B7, with statistical differences assessed using paired t-tests at p < 0.05. The results demonstrate that segmentation-guided ROI extraction combined with deep feature-based classification and weighted probabilistic fusion provides an effective approach for automated brain tumor analysis.