Adaptive example selection for prototype based explainable mitosis detection in digital pathology.

Area of Science:

• Artificial Intelligence
• Computational Pathology
• Medical Image Analysis

Background:

• Black-box neural networks pose challenges for AI safety in critical medical applications like histopathology.
• Interpretability is essential for clinicians to trust and effectively utilize AI tools in cancer diagnostics.

Purpose of the Study:

• To introduce Adaptive Example Selection (AES), a novel explainable AI framework for enhancing the interpretability of deep learning models in mitosis detection.
• To enable clinicians to visualize AI reasoning, assess uncertainty, and perform contrastive analyses for improved diagnostic confidence.

Main Methods:

• Developed a prototype-based explainable AI framework (AES) that retrieves supporting and contradicting image prototypes.
• Integrated AES with a Faster R-CNN detector for robust mitosis detection and cross-tumor performance evaluation.
• AES approximates the model's confidence surface locally to generate case-specific explanations linked to interpretable exemplars.

Main Results:

• The Faster R-CNN detector achieved strong cross-tumor performance, with an F1-score of 0.84 on the Canine Cutaneous Mast Cell Tumor dataset.
• AES provided concise explanations that accurately captured local decision boundaries and linked predictions to relevant prototypes.
• Demonstrated how similarity to mitotic and non-mitotic prototypes influences graded confidence, offering a nuanced view beyond discrete predictions.

Conclusions:

• AES significantly enhances the transparency and trustworthiness of AI-assisted mitosis detection systems.
• The framework facilitates practical adoption of AI in cancer diagnostics by enabling clinicians to understand and validate model predictions.
• AES represents a step forward in making AI decision-making in histopathology more accessible and reliable for clinical use.