Grad-CAM Sensitivity to Orientation Transformations in Breast Cancer Histopathological Images

Breast cancer is the second most prevalent among women, and early diagnosis is essential to reduce mortality. Machine learning models are being used to improve diagnostic accuracy, especially through biopsies. However, the growing complexity of neural networks has raised concerns about result interpretability, driving the adoption of explainability methods such as the Grad-CAM, which generates saliency maps in CNN models. This paper investigates the sensitivity of Grad-CAM to changes in orientation and position in breast histopathological images. CNN's ResNet50, VGG16, and Xception were evaluated, trained with the BreakHis dataset, which comprises images at different magnifications. Test samples were modified by rotations and flips to assess the robustness of both the models and the explanations. Xception achieved the best results in terms of accuracy (94.47%) and F1-score (95.85%) with 100X images, although the explanatory maps varied according to the applied transformation.