Towards Robust Breast Cancer Diagnosis: A Hybrid Deep Learning Ensemble Framework

Breast cancer remains a leading cause of female mortality worldwide, highlighting the critical need for precise diagnosis to improve patient outcomes. Accurately distinguishing cancerous cells from healthy tissue in microscopy images is challenging due to tumor complexity. While deep learning (DL) models show promise for classification, further advancements are essential to enhance accuracy. This study introduces ResEff-NET, a novel deep stacked ensemble framework designed to improve breast cancer cell classification. ResEff-NET integrates the pre-trained EfficientNetB1 and ResNet101 models: first, by augmenting each base classifier with additional layers to boost feature extraction; second, by employing a stacking technique that combines their predictions to achieve significantly higher accuracy than individual models. Evaluated on a dataset of 21,848 labeled breast cell microscopy images, ResEff-NET demonstrated superior performance through comparative analysis with its backbone models and existing methods, supplemented by Grad-CAM and feature map analysis. Experimental results confirm the model’s stability, effectiveness, and substantial improvement in accuracy. These findings underscore ResEff-NET’s potential as a valuable tool to assist medical professionals in breast cancer diagnosis and decision-making.