Deep Ensemble Model for Fine-Grained Breast Tumor Classification Across Multi-Magnification Samples

Breast cancer is a type of malignant tumor that originates in the cells of the breast tissue. Breast cancer remains one of the leading causes of cancer-related mortality among women worldwide, and histopathological analysis plays a crucial role in its diagnosis and subtyping. Despite advancements in automated classification, the heterogeneity of tissue morphology across varying magnification levels remains a challenge to conventional deep learning approaches. In this study, we present a robust and interpretable multiclass classification framework for breast cancer tumor subtypes using the Breast Cancer Histopathological Image Classification (BreaKHis) dataset. To address the challenges of tissue heterogeneity, we introduce a novel ensemble architecture that integrates complementary features extracted from three state-of-the-art Convolutional Neural Networks (CNNs) (ResNet50, VGG16, and EfficientNet-B0) via an adaptive fusion strategy. The proposed model demonstrates superior generalization across all four magnification levels ($40 \times, 100 \times,\ 200 \times$, and $400 \times$), achieving an average classification accuracy of 99.4%, a macro-averaged F 1-score of 97.2% and low inter-fold variance, validated through stratified 5-fold cross-validation. This comprehensive pipeline not only outperforms several state-of-theart and baseline models but also advances explainable neural networks in histopathology, offering a transparent and reliable computational proof of concept for future decision-support tools in precision oncology.