Breast cancer is the second leading cause of cancer-related deaths among women, following lung cancer, as of 2024. Conventional cancer diagnosis relies on the manual examination of biopsied tissues by pathologists, a time-consuming process that may vary based on individual expertise. Early detection and accurate diagnosis are crucial for effective treatment planning and patient care. The advent of whole-slide scanners has revolutionized this process by enabling the use of Computer-Aided Detection (CAD) systems for automated analysis. In this study, we utilize state-of-the-art Convolutional Neural Networks (CNNs), specifically EfficientNetV1 and EfficientNetV2, for the binary classification of the BreakHis dataset-a collection of histopathological images categorized as benign or malignant breast tissues. To address the challenge of limited annotated data, we apply data augmentation and transfer learning techniques. Model interpretability is enhanced using the Grad-CAM technique, which generates localization maps highlighting critical regions relevant to predictions. Furthermore, ensemble learning is employed to improve classification performance. We utilize unweighted averaging and majority voting to combine predictions from multiple trained models. Additionally, we propose two ensemble architectures that integrate different trained EfficientNet models. Our framework achieves a classification accuracy of 99.58%, outperforming conventional CNN models on the BreakHis dataset. This study highlights the potential of ensemble learning to enhance diagnostic accuracy in breast cancer detection.