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

The article presents the People-Centred Medical Image Analysis (PecMan) framework, which integrates fairness-aware human-AI cooperation for medical image classification. PecMan employs subgroup-specialized predictors with a dynamic gating mechanism to allocate decisions between automated systems and human experts without requiring sensitive attributes during testing. The introduction of the FairHAI benchmark allows for the evaluation of predictive accuracy, subgroup equity, and human involvement, with experimental results showing that PecMan outperforms traditional methods that treat fairness and human-AI collaboration as separate challenges.