<![CDATA[Class imbalance in breast cancer imaging often leads to models prioritizing the majority class, reducing sensitivity to actual cancer cases. This study evaluates data augmentation as a class balancing strategy for breast cancer classification using VGG19 with transfer learning. The model was trained and tested in two settings: before and after augmentation, to measure performance improvement. The results show a clear improvement after balancing, with accuracy rising from 94.63% to 97.59%, recall and specificity increasing from about 85.60% to 97.58%, and the F1 score rising from 0.8933 to 0.9759, indicating better balance between precision and recall. Interpretability analysis using Grad-CAM supports this improvement, with activations before augmentation being spread out and sometimes focusing on background artifacts, while the heatmap after augmentation concentrated on the lesion region, indicating that the network learned clinically meaningful features. Overall, the findings demonstrate that targeted augmentation effectively addresses class imbalance, enhances generalization, and improves lesion detection with VGG19. This approach enhances cancer sensitivity while reducing false alarms, supporting its potential for adoption in computer-aided diagnostic pipelines to provide more reliable breast cancer detection in clinical practice.]]>
Copyrights © 2025
