People-Centred Medical Image Analysis
New human-AI gating system cuts bias while respecting clinician workload limits.
A new paper on arXiv (2604.26991) introduces PecMan (People-Centred Medical Image Analysis), a human-AI framework designed to address the limited clinical adoption of medical imaging AI. The authors identify two key barriers: performance biases across diverse patient populations and poor workflow integration that disrupts clinical routines. Prior work on workflow integration (e.g., Learning to Defer and Learning to Complement) and AI fairness has treated these challenges separately, overlooking their interdependence and real-world constraints like limited clinician availability.
PecMan uses a dynamic gating mechanism that assigns each case to AI, a clinician, or both, respecting clinician workload limits while jointly optimizing fairness, accuracy, and workflow effectiveness. The team also introduces the FairHAI benchmark for evaluating trade-offs between these factors. Experiments on this benchmark show PecMan consistently outperforms existing methods, offering a pathway to more trustworthy and clinically viable AI systems. Code will be released upon paper acceptance.
- PecMan uses a dynamic gating mechanism to assign cases to AI, clinicians, or both under workload constraints.
- The new FairHAI benchmark evaluates trade-offs between accuracy, fairness, and clinician workload.
- PecMan outperforms existing methods (L2D, L2C) in joint optimization of fairness, accuracy, and workflow.
Why It Matters
Medical AI that balances fairness and workflow could finally gain clinician trust and regulatory approval.