A ResNet-50 Classifier with Grad-CAM and SHAP for Explainable Breast Cancer Detection in Dense Mammography: Comparison with Human Radiologist Performance

Technical Summary: A ResNet-50 Classifier with Grad-CAM and SHAP for Explainable Breast Cancer Detection in Dense Mammography

Problem Statement
Breast cancer remains the most frequently diagnosed malignancy among women globally. While screening mammography reduces mortality, its efficacy is significantly compromised in women with dense breast tissue, where overlapping fibroglandular structures obscure lesions. Radiologists reading dense mammograms achieve a sensitivity of only 60–70%, leading to missed cancers and delayed diagnoses. Although deep learning models have demonstrated high accuracy in retrospective studies, they typically function as "black boxes," lacking the transparency required for clinical adoption. Physicians cannot verify the model's reasoning or identify potential biases, limiting trust and integration into workflows. There is an urgent need for diagnostic systems that not only achieve high accuracy on dense mammograms but also provide interpretable, clinically actionable explanations.

Methodology
The study employed a supervised deep learning approach using a ResNet-50 architecture trained on a balanced dataset derived from the MammoBench v2 collection. This dataset aggregates six mammography sources (INbreast, CMMD, DDSM, DMID, KAU-BCMD, and CDD-CESM), totaling 19,731 images. To address class imbalance, the authors constructed a balanced dataset of 16,368 images (8,184 malignant and 8,184 non-cancer), which was split into training (60%), validation (20%), and test (20%) sets.

The model architecture utilized a ResNet-50 backbone pretrained on ImageNet, with the top classification layer removed and replaced by a global average pooling layer, dropout layers, and dense layers for binary classification. The model was trained using binary cross-entropy loss with the Adam optimizer, employing early stopping and learning rate reduction to prevent overfitting.

To address the interpretability gap, the study integrated two Explainable AI (XAI) techniques:

1. Grad-CAM (Gradient-weighted Class Activation Mapping): Applied to the conv5_block3_conv layer to generate spatial heatmaps. This visualizes the specific regions of the mammogram that most influenced the model's prediction, answering the "where" question.
2. SHAP (SHapley Additive exPlanations): Applied to a stratified sample of 300 test images using KernelSHAP. This method quantifies the contribution of individual features to the model's output, providing global interpretability and assessing prediction confidence.

Key Contributions
The paper makes four primary contributions:

1. Performance Benchmarking: It demonstrates that a standard ResNet-50 classifier can achieve higher sensitivity (72.2%) than the reported human range (60–70%) for dense breast mammography, while maintaining competitive specificity.
2. Comprehensive XAI Framework: It combines spatial explanations (Grad-CAM) with quantitative feature importance (SHAP) to generate clinically actionable confidence thresholds.
3. Error Characterization: It utilizes Grad-CAM visualizations to distinguish between error types. False negatives are characterized by weak, diffuse activation failing to localize the lesion, whereas false positives exhibit either strong focal activation over benign masses (misinterpretation) or diffuse activation over dense parenchyma (over-interpretation).
4. Clinical Deployment Guidelines: It derives specific trust guidelines based on SHAP analysis, recommending the model's use as a supportive second reader rather than an autonomous screener.

Results
On a held-out test set of 3,274 balanced images, the model achieved:

• Accuracy: 76.45%
• Sensitivity (Recall): 72.2% (detecting 1,182 of 1,637 malignant cases)
• Specificity: 80.7% (correctly identifying 1,321 of 1,637 non-cancer cases)
• AUC-ROC: 0.846
• Average Precision: 0.835

The sensitivity of 72.2% exceeds the typical human performance range for dense breasts, though the specificity is modestly lower than the radiologist norm of 85–90%. The authors note this trade-off is clinically acceptable for a second-reader tool where missing a cancer is the primary risk.

SHAP analysis of 300 randomly sampled test images revealed that all predictions fell within an "uncertain" probability range (0.3–0.7), with a mean probability of 0.351 and a mean entropy of 0.64. This indicates the model expresses appropriate caution on unselected mammograms and defaults to baseline prevalence in the absence of strong evidence.

Significance and Claims
The paper claims that a standard deep learning classifier, when paired with robust explainability methods, can surpass human sensitivity in the challenging domain of dense breast tissue while providing transparent reasoning. The study emphasizes that the model is best deployed as a second reader, not an autonomous diagnostic tool.

Based on the SHAP findings, the authors propose three specific clinical recommendations:

1. Trust the model when output probabilities are below 0.3 or above 0.7, as these represent high-confidence predictions.
2. Require mandatory radiologist review for predictions between 0.4 and 0.6, where model uncertainty is highest.
3. Deploy the model to flag high-confidence cases and defer ambiguous ones to human experts, thereby enhancing rather than replacing clinical judgment.

The authors conclude that this framework offers a practical pathway for integrating explainable AI into breast cancer screening, potentially improving early detection rates and reducing missed cancers in dense-breast populations. The study acknowledges limitations, including the lack of a matched reader study, the use of a balanced dataset that overrepresents cancer prevalence, and the absence of demographic annotations for subgroup bias analysis.