Prototype-Based Interpretable Breast Cancer Prediction Models: Analysis and Challenges

Shreyasi Pathak, Jörg Schlötterer, Jeroen Veltman, Jeroen Geerdink, Maurice van Keulen, Christin Seifert

Abstract

Deep learning models for breast cancer detection often operate as "black boxes," hindering clinical trust. Prototype-based models aim to provide interpretability by design, yet their evaluation often relies on anecdotal evidence. This paper introduces the Prototype Evaluation Framework for Coherence (PEF-Coh) to systematically measure the medical relevance and quality of these explanations. Our analysis reveals that while interpretable models achieve competitive diagnostic performance, a significant portion of their reasoning relies on medically irrelevant features, underscoring the need for better alignment with clinical domain knowledge.

Introduction

The integration of AI into mammography screening holds potential for improving early detection. However, the lack of transparency in standard deep learning models remains a barrier. Prototype-based models (e.g., ProtoPNet) address this by making predictions based on the similarity of input image patches to learned "prototypes"—representative features of specific abnormalities. Despite their promise, there has been no rigorous method to evaluate whether these prototypes truly capture medical concepts or merely exploit background noise.

Methodology: The PEF-Coh Framework

To move beyond qualitative "cherry-picking," the researchers introduced the PEF-Coh framework. This framework evaluates prototype quality across seven quantitative dimensions grounded in medical domain knowledge, as detailed in the table below.

Table 1: The PEF-Coh framework metrics for evaluating prototype coherence.

The metrics include:

  • Compactness: Measures the size of the explanation (number of prototypes).
  • Relevance: Quantifies the overlap between prototypes and annotated Regions of Interest (ROIs).
  • Specialization: Assesses if prototypes represent single, pure medical concepts.
  • Uniqueness & Coverage: Evaluate the diversity and breadth of medical features learned.
  • Localization: Measures the accuracy of prototype activation maps against ground-truth lesions.
  • Class-specificity: Ensures prototypes are correctly associated with benign or malignant outcomes.

The study evaluated three state-of-the-art models—ProtoPNet, BRAIxProtoPNet++, and PIP-Net—using the CBIS-DDSM, VinDr-Mammo, and CMMD datasets.

Key Insights and Results

The Relevance Gap

A critical finding is the low relevance of learned prototypes. Only 28-36% of prototypes across models actually pointed to lesions. Many were triggered by non-diagnostic features like pectoral muscles or image background. This suggests that "interpretable" models may still be learning spurious correlations rather than medical features.

Performance vs. Interpretability

Interpretable models demonstrated diagnostic performance comparable to black-box architectures like ConvNext. For instance, BRAIxProtoPNet++ achieved an AUC of 0.86 on the CMMD dataset. However, a trade-off exists between simplicity and depth: PIP-Net was the most compact but lacked the "Coverage" to recognize a wide variety of medical abnormalities.

Localization and Redundancy

Prototype-based models outperformed standard saliency map techniques (e.g., GMIC) in lesion localization. However, they frequently suffered from redundancy, learning multiple near-identical prototypes for the same feature, which complicates human interpretation without adding diagnostic value.

Conclusion and Future Directions

While prototype-based models are a step toward transparent AI, current training methods fail to ensure medical coherence. The authors propose a "human-in-the-loop" workflow where clinicians verify and name global prototypes before deployment. This ensures the model's reasoning is grounded in the medical lexicon, transforming AI from a black box into a collaborative clinical tool.

Summary updated

BibTeX
@inproceedings{pathak2024prototype,
  author = {Shreyasi Pathak and Jörg Schlötterer and Jeroen Veltman and Jeroen Geerdink and Maurice van Keulen and Christin Seifert},
  title = {{Prototype-Based Interpretable Breast Cancer Prediction Models: Analysis and Challenges}},
  booktitle = {xAI},
  year = {2024},
  url = {https://openreview.net/forum?id=zyqLaQS0S4},
}
Figure