MPD$^2$-Router: Mask-aware Multi-expert Prior-regularized Dual-head Deferral Router in Glaucoma Screening and Diagnosis
MPD²-Router is a machine learning framework that improves glaucoma screening by intelligently routing difficult cases between AI systems and human experts based on availability, uncertainty, and image quality. The system achieves better clinical outcomes than AI-alone approaches while maintaining balanced expert utilization across multiple international datasets.
MPD²-Router addresses a critical gap in medical AI deployment: the challenge of effectively triaging cases between automated systems and human experts in ophthalmology. The framework moves beyond simple defer-or-diagnose decisions to create a constrained routing system that considers real-world operational constraints like expert availability, varying reader behavior, and workload distribution. This represents a maturation in learning-to-defer research, acknowledging that clinical AI systems don't operate in isolation but within complex healthcare workflows.
The technical innovation combines multiple signal sources—morphological features, image quality assessment, uncertainty estimation, and out-of-distribution detection—through a dual-head policy mechanism with mask-aware gating. The asymmetric cost-sensitive training objective with augmented-Lagrangian constraints directly models the asymmetric clinical harm where missed glaucoma cases carry higher cost than false positives. The rank-majorization regularizer prevents common failure modes like expert collapse, ensuring sustainable expert utilization without artificially forcing uniform allocation.
For healthcare AI development, this work demonstrates how production deployment requirements reshape model architecture. The system achieves Pareto-optimality across multiple metrics (F1, MCC, cost) while maintaining robustness under cross-domain shift, critical for deploying screening tools across diverse clinical settings and populations. The evaluation across three international cohorts (REFUGE, CHAKSU, ORIGA) provides evidence of generalization beyond single-population training.
The framework's ability to lower clinical cost while improving diagnostic accuracy at moderate deferral rates addresses the economic and safety pressures facing healthcare systems. Future development likely focuses on personalizing expert routing based on individual reader expertise patterns and extending the approach to other screening domains facing similar human-AI coordination challenges.
- →MPD²-Router routes glaucoma cases between AI and human experts based on availability, uncertainty, and image quality rather than simple binary defer decisions
- →The system achieves Pareto-optimal performance across F1, MCC, and clinical cost metrics while maintaining balanced expert utilization
- →Asymmetric cost-sensitive training models real clinical harm where missed glaucoma diagnosis carries higher penalty than false positives
- →Framework demonstrates cross-domain robustness across three international datasets despite training on REFUGE cohort alone
- →Addresses healthcare deployment reality where expert availability and workload constraints shape optimal AI-human collaboration strategies