Pathologist-like explainable AI for interpretable Gleason grading in prostate cancer

Explainability improves both trust and performance: The AI system trained to explain its reasoning using pathologists' own terminology achieved better accuracy (Dice score of 0.713) than systems trained only to predict outcomes (0.691), demonstrating that making AI decisions transparent doesn't require sacrificing performance—it can actually enhance it.

Human disagreement is a feature, not a bug: By incorporating "soft labels" that capture uncertainty from the 54 pathologists' varied annotations, the system acknowledged that even experts sometimes disagree on cancer grading. This approach yielded stronger results than treating one annotation as definitively correct, suggesting responsible AI should embrace rather than hide human subjectivity in medical decisions.

Pathologist-centered design builds appropriate trust: Rather than using post-hoc explanation methods (which add explanations after the fact), the system was built from the ground up to communicate in pathologists' own language and reasoning patterns, making it more likely that medical professionals would understand and appropriately trust the AI's recommendations.

Dataset transparency enables accountability: By releasing a dataset of 1,015 annotated tissue images with detailed expert reasoning, the researchers created a foundation for ongoing scrutiny and improvement, allowing other researchers to validate the approach and understand how the AI was trained—essential for responsible deployment in clinical settings.