Enhanced Interpretability in Breast Cancer Detection: Combining Grad‐ CAM With Selective Layer Freezing in Deep Learning

ABSTRACT This study aims to develop a novel deep learning‐based approach that integrates selective layer freezing, cyclic learning rate scheduling, and Grad‐CAM visualization to address the challenges of class imbalance, limited interpretability, and adaptability in breast cancer detection from mammographic images. The proposed framework utilized ResNet50 and VGG19 architectures, fine‐tuned with selective layer freezing to optimize the balance between general feature preservation and domain‐specific adaptation. Mammographic images comprising 8398 images (4194 malignant and 4204 benign) were preprocessed using resizing, histogram equalization, normalization, and data augmentation to enhance feature extraction and mitigate class imbalance. The dataset was divided into training, validation, and test sets (80:15:5), with an additional 136 external mammograms included for validation. Grad‐CAM was applied to provide visual interpretability by highlighting diagnostic regions such as abnormal masses and architectural distortions. Performance was evaluated using metrics such as accuracy, precision, recall, F1‐score, and AUC. The ResNet50 model achieved an AUC of 0.97 across all freezing ratios, with the 50% freezing ratio delivering the most balanced performance (accuracy: 97%, precision: 97%, recall: 97%). In comparison, the VGG19 model achieved a maximum AUC of 0.95 at the 50% freezing ratio. Grad‐CAM outputs confirmed the interpretability of the models, with sharp and clinically relevant visualizations provided by ResNet50. External validation further demonstrated the robustness and generalizability of the proposed framework. The proposed framework effectively combines high diagnostic accuracy with enhanced interpretability, making it a valuable tool for breast cancer detection. Future work will focus on multi‐class classification and large‐scale clinical validation.