People-Centred Medical Image Analysis via Fairness-Aware Human-AI Cooperation

AI systems in medical imaging perform differently across patient groups: The paper demonstrates that machine learning models show "subgroup-dependent performance," meaning they work better for some patient populations than others—a critical fairness issue when these systems help doctors make diagnostic decisions.

Combining human expertise with AI strategically improves fairness: Rather than choosing between fully automated decisions or full human review, PecMan dynamically routes cases to AI alone, human experts alone, or both working together based on confidence levels. This approach achieved "consistently improved trade-offs" between accuracy and fairness compared to methods addressing these problems separately.

Fairness and efficiency create real trade-offs that need transparent evaluation: The paper introduces the FairHAI benchmark to measure three competing goals simultaneously: predictive accuracy, equitable performance across patient groups, and minimizing unnecessary human involvement. This highlights that improving fairness often requires accepting some efficiency costs—a reality that should inform policy decisions.

Sensitive patient information doesn't need to be visible at deployment: PecMan's system works without requiring race, gender, or other sensitive attributes during actual clinical use, addressing privacy concerns while still achieving fairness by learning subgroup-specific patterns during training.

One-size-fits-all AI deployment in healthcare may be inadequate: The research suggests that responsible medical AI requires systems sophisticated enough to recognize when they're uncertain about specific patient populations and route those cases appropriately—rather than applying uniform decision rules across all patients.