Towards Trustworthy Breast Cancer Diagnosis: A Comparative Explainability Stability Study of DenseNet121 and Vision Transformers

Technical Summary: Towards Trustworthy Breast Cancer Diagnosis

Problem Statement
Deep learning models have demonstrated high success in medical image classification, particularly for breast cancer diagnosis using histopathology images. However, these models often function as "black boxes," lacking interpretability for clinicians, which undermines trust in safety-critical applications. While Explainable Artificial Intelligence (XAI) tools like SHapley Additive exPlanations (SHAP) provide visual rationales for predictions, there is a significant gap in understanding the stability and robustness of these explanations. Specifically, it remains unclear how explanation stability differs between convolutional architectures (e.g., DenseNet) and transformer-based architectures (e.g., Vision Transformers) when input data undergoes minor perturbations such as noise, brightness shifts, or contrast scaling. Without stable explanations, clinicians may lose confidence in automated diagnostic systems even if the classification accuracy remains high.

Methodology
This study proposes a comparative framework to evaluate the explainability stability of DenseNet121 (a Convolutional Neural Network) and Vision Transformer (ViT) models on the Breast Histopathology Images dataset. The methodology ensures a fair comparison by training both architectures with identical preprocessing, optimization parameters, and perturbation settings.

1. Data Preparation: The dataset was stratified into training, validation, and test sets. Images were resized to 224x224, normalized using ImageNet statistics, and augmented with random horizontal flipping.
2. Model Architecture:
   • DenseNet121: Utilizes dense connectivity to facilitate feature reuse and gradient propagation, focusing on local spatial contexts.
   • Vision Transformer (ViT-B/16): Divides images into patches and utilizes self-attention mechanisms to model long-range dependencies and global context.
3. Explainability Generation: SHAP (specifically GradientExplainer) was employed to generate visual attribution maps for both models, highlighting regions driving the classification decision.
4. Stability Analysis: To assess robustness, input images were subjected to three types of perturbations: Gaussian noise, brightness shift, and contrast scaling.
5. Quantitative Metrics: The stability of the explanations was measured using:
   • Cosine Similarity: To measure directional consistency between original and perturbed explanations.
   • Mean Absolute Difference (MAD): To quantify the magnitude of changes in attribution values.
   • Structural Similarity Index Measure (SSIM): To evaluate perceptual similarity.
6. Statistical Validation: Paired t-tests and Cohen's d effect size analysis were conducted to determine if observed differences in stability between the two architectures were statistically significant.

Key Contributions

• Comparative Stability Study: This work fills a gap in the literature by directly comparing the explanation stability of CNN-based and Transformer-based architectures in the context of breast histopathology, a domain where such comparisons are scarce.
• Dual-Metric Evaluation: The study moves beyond predictive accuracy to simultaneously evaluate classification performance and the robustness of XAI outputs under input perturbations.
• Statistical Rigor: Unlike previous studies that may rely on visual inspection or single-metric comparisons, this research employs rigorous statistical testing (t-tests and effect size analysis) to validate the significance of stability differences.
• Trade-off Identification: The paper explicitly identifies and quantifies the trade-off between high classification accuracy/localized explanations (DenseNet) and explanation consistency under perturbation (ViT).

Experimental Results

• Classification Performance: DenseNet121 outperformed the Vision Transformer in predictive metrics, achieving 91% accuracy and a 0.965 ROC-AUC, compared to the ViT's 89% accuracy and 0.948 ROC-AUC. DenseNet121 also demonstrated higher sensitivity and recall for malignant samples, which is critical for minimizing false negatives in medical diagnosis.
• Explanation Localization: DenseNet121 produced more spatially coherent and localized attribution maps, aligning closely with diagnostically relevant tissue structures. In contrast, ViT explanations were more distributed, reflecting its global attention mechanism.
• Stability Under Perturbation: The Vision Transformer demonstrated superior explanation stability. It achieved higher Cosine Similarity (0.261 vs. 0.218) and SSIM (0.667 vs. 0.509) values, and a lower MAD (0.00044 vs. 0.00138) compared to DenseNet121. This indicates that ViT explanations remained more consistent when the input images were altered by noise or contrast changes.
• Statistical Significance: The difference in explanation stability between the two models was found to be statistically significant (p < 0.001), with a Cohen's d effect size of -0.341, confirming that the observed stability advantage of the ViT is not due to chance.

Significance and Claims
The paper argues that trustworthy medical AI systems must be evaluated on both classification accuracy and the reliability of their explanations. The findings highlight a critical trade-off: while DenseNet121 offers superior diagnostic accuracy and localized interpretability, the Vision Transformer provides more robust and consistent explanations in the presence of input variations.

The authors conclude that for clinical decision support, relying solely on accuracy is insufficient. Future medical AI frameworks must incorporate the stability of explanations as a core metric. The study suggests that the choice of architecture should depend on the specific clinical requirement: if the priority is maximum diagnostic sensitivity, DenseNet121 is preferable; if the priority is maintaining consistent interpretability despite minor imaging inconsistencies (common in real-world scenarios), the Vision Transformer offers a distinct advantage. The work underscores the need for a holistic assessment of AI systems to ensure they are not only accurate but also clinically trusted and reliable.