MPD$^2$-Router: AI Defers Glaucoma Cases to Human Experts
Researchers have unveiled an innovative glaucoma screening system called MPD$^2$-Router. This framework reimagines how we approach ophthalmic triage by using a controlled human-AI routing method to decide on case deferrals and which specialists to consult. It features a dual-head strategy for deferral and allocation, utilizing mask-aware Gumbel-sigmoid gating to adapt to specific cases while considering factors like uncertainty, morphology, image quality, and out-of-distribution data. The training employs a unique cost-sensitive objective with an augmented-Lagrangian deferral budget and a group-specific distribution prior to prevent expert collapse. This system addresses issues like expert availability, different reader behaviors, workload balance, and case complexity. The study can be found on arXiv under ID 2605.08024.
Key facts
- MPD$^2$-Router is a mask-aware multi-expert deferral framework for glaucoma screening.
- It recasts ophthalmic triage as constrained human-AI routing.
- The system decides whether to defer and to which available expert.
- It uses a dual-head deferral/allocation policy with mask-aware Gumbel-sigmoid gating.
- Training uses an asymmetric cost-sensitive objective with augmented-Lagrangian deferral budget.
- A group-specific distribution prior and rank-majorization JS regularizer prevent expert collapse.
- The framework addresses expert availability, heterogeneous reader behavior, workload imbalance, asymmetric diagnostic harm, case difficulty, and deployment shift.
- Published on arXiv with ID 2605.08024.
Entities
Institutions
- arXiv