MPD$^2$-Router: Mask-aware Multi-expert Prior-regularized Dual-head Deferral Router in Glaucoma Screening and Diagnosis

TL;DR

MPD$^2$-Router is a novel mask-aware multi-expert routing framework for glaucoma screening that improves clinical cost and accuracy by effectively deferring uncertain cases to human experts while considering expert availability and workload.

Contribution

It introduces a constrained human-AI routing approach with a dual-head policy and mask-aware gating, addressing expert availability, workload, and heterogeneity in glaucoma diagnosis.

Findings

Reduces clinical cost and improves MCC over AI-only methods.

Maintains balanced expert utilization and robustness under domain shifts.

Achieves Pareto optimality in F1, MCC, and cost metrics.

Abstract

Learning-to-defer (L2D) can make glaucoma screening safer by routing difficult/uncertain cases to humans, yet standard formulations overlook expert availability, heterogeneous readers behavior, workload imbalance, asymmetric diagnostic harm, case difficulty from morphology and deployment shift. We introduce MPD$^2$-Router, a mask-aware multi-expert deferral framework that recasts ophthalmic triage as constrained human--AI routing: whether to defer and to which available expert. It couples a dual-head deferral/allocation policy with mask-aware Gumbel--sigmoid gating that strictly enforces per-sample availability, and fuses uncertainty, morphology, image-quality, and OOD signals. Training uses an asymmetric cost-sensitive objective with an augmented-Lagrangian deferral budget, a group-specific distribution prior, and a rank-majorization JS regularizer that jointly prevent expert collapse without forcing uniform allocation. Across three cross-national glaucoma cohorts (REFUGE, CHAKSU, ORIGA) with a frozen REFUGE-trained backbone, MPD$^2$-Router substantially lowers clinical cost and improves MCC over AI-only at a moderate deferral rate. It is Pareto-optimal in F1--MCC--cost, robust under cross-domain shift, and yields balanced expert utilization.