<![CDATA[Breast cancer remains one of the leading causes of cancer-related mortality among women worldwide, highlighting the need for diagnostic systems that are both accurate and interpretable. Although transfer learning has shown promising results in histopathological image classification, studies simultaneously examining predictive performance, statistical reliability, and interpretability remain limited. This study proposes an explainable transfer learning framework for breast cancer histopathology classification and investigates the relationship between classification performance and visual interpretability. Experiments were conducted using 2,013 histopathological images from the BreakHis dataset at 200× magnification. Three pretrained architectures, ResNet50, DenseNet121, and EfficientNetB0, were trained and evaluated under identical preprocessing, augmentation, and training settings. Performance was assessed using accuracy, precision, recall, F1-score, AUC, confidence intervals, McNemar testing, confusion matrix analysis, and Grad-CAM visualization. Results showed that DenseNet121 achieved the most balanced classification performance and the highest discriminative capability among the evaluated models. Statistical analysis confirmed significant performance differences, while Grad-CAM visualizations demonstrated more focused and diagnostically relevant activation regions. These findings suggest that models learning more discriminative histopathological representations tend to generate more meaningful visual explanations. The study emphasizes integrating predictive performance, statistical validation, and explainability to support reliable and transparent artificial intelligence systems for breast cancer diagnosis.]]>
