Explainable AI for computational pathology identifies model limitations and tissue biomarkers.

Area of Science:

• Computational pathology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning models in digital pathology lack transparency, hindering clinical trust and adoption.
• Explainable AI (XAI) is crucial for enhancing model reliability and understanding decision-making processes.

Purpose of the Study:

• To introduce HIPPO, an XAI framework for digital pathology.
• To enable quantitative hypothesis testing, bias detection, and model evaluation using image counterfactuals.
• To improve the trustworthiness and clinical adoption of AI models in pathology.

Main Methods:

• HIPPO systematically modifies tissue regions in whole slide images to create counterfactuals.
• The framework was applied to breast metastasis detection, cancer prognostication, and glioma mutation classification.
• HIPPO's performance was compared against traditional metrics and attention-based XAI methods.

Main Results:

• HIPPO identified critical model limitations in metastasis detection missed by standard metrics and attention.
• For prognostication, HIPPO provided more nuanced insights into tissue drivers than attention.
• HIPPO aided hypothesis generation for melanoma immunotherapy and improved identification of false negatives in mutation classification.

Conclusions:

• HIPPO expands the XAI toolkit for computational pathology, offering deeper insights into model behavior.
• The framework supports the development, deployment, and regulation of trustworthy AI in digital pathology.
• HIPPO promotes broader adoption of AI in clinical and research settings by enhancing model transparency.