GRAPHITE: Graph-based interpretable tissue examination for enhanced explainability in breast cancer histopathology.

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in oncology

Background:

• Deep learning models in cancer diagnosis lack interpretability, hindering clinical trust and adoption.
• Explainable AI (XAI) is crucial for understanding black-box models in medical histopathology.
• Breast cancer tissue microarray (TMA) analysis requires interpretable AI tools for accurate diagnosis.

Purpose of the Study:

• To introduce GRAPHITE (Graph-based Interpretable Tissue Examination), a post-hoc explainable framework for breast cancer TMA analysis.
• To enhance the clinical trustworthiness and interpretability of deep learning models in histopathology.
• To provide visualisations that align with pathologist diagnostic reasoning.

Main Methods:

• GRAPHITE utilizes a multiscale approach, extracting patches at various magnification levels.
• An hierarchical graph is constructed, employing graph attention networks (GAT) with scalewise attention (SAN).
• The model was trained on 140 tumour TMA cores and 140 benign samples, tested on 53 annotated TMA samples.

Main Results:

• GRAPHITE achieved a mean average precision (mAP) of 0.56 and an AUROC of 0.94.
• The framework demonstrated high threshold robustness (ThR) of 0.70.
• GRAPHITE achieved the highest area under the decision curve (AUDC) of 4.17e+5, indicating reliable clinical decision support.

Conclusions:

• GRAPHITE offers a clinically valuable tool for computational pathology in breast cancer diagnosis.
• The framework provides interpretable visualisations that support pathologist diagnostic reasoning.
• GRAPHITE has the potential to advance precision medicine through enhanced AI interpretability.