Back to AI Research

AI Research

MPD$^2$-Router: Mask-aware Multi-expert Prior-regul... | AI Research

Key Takeaways

  • MPD$^2$-Router: Mask-aware Multi-expert Prior-regularized Dual-head Deferral Router in Glaucoma Screening and Diagnosis This paper introduces a new framework...
  • 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.
  • 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.
Paper AbstractExpand

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.

MPD$^2$-Router: Mask-aware Multi-expert Prior-regularized Dual-head Deferral Router in Glaucoma Screening and Diagnosis
This paper introduces a new framework designed to make AI-assisted glaucoma screening safer and more efficient. While many AI systems are designed to provide a single diagnosis, real-world clinical practice requires a more nuanced approach: determining when an AI can act autonomously and when a case should be referred to a human expert. The MPD$^2$-Router addresses this by treating screening as a "routing" problem, ensuring that difficult or uncertain cases are sent to the most appropriate available human expert based on their specific skills and current workload.

A More Realistic Approach to Triage

Standard AI models often treat human experts as interchangeable, "perfect" sources of truth. However, in reality, ophthalmologists and graders have different levels of experience and specialties. Furthermore, AI systems often struggle with "out-of-distribution" data, such as low-quality images or unusual eye structures. The MPD$^2$-Router moves beyond simple classification by incorporating multiple signals—including AI uncertainty, image quality, and optic-disc morphology—to decide whether to trust the AI or escalate the case to a human.

How the Routing System Works

The system uses a "dual-head" architecture to manage the decision-making process. The first head decides whether a case needs human intervention, while the second head determines which specific expert should handle it. To ensure this works in practice, the model uses "mask-aware" technology, which strictly enforces that a case is only routed to an expert who is actually available and capable of handling it. By using a specialized training method, the system prevents "expert collapse," where all difficult cases are unfairly dumped onto a single, overburdened specialist.

Balancing Clinical Costs and Expert Workload

A key challenge in medical AI is balancing the need for accuracy with the reality of limited human resources. The MPD$^2$-Router is trained using an "asymmetric cost-sensitive" objective, which acknowledges that missing a glaucoma diagnosis is more harmful than a false alarm. It also uses a hierarchical prior to distribute the workload across a team of experts based on their reliability and availability. This ensures that the system remains useful and fair, preventing any single expert from being overwhelmed while maintaining high diagnostic standards.

Performance and Robustness

The researchers tested the MPD$^2$-Router across three different international datasets (REFUGE, CHAKSU, and ORIGA). The results show that the framework significantly lowers clinical costs and improves diagnostic accuracy compared to using AI alone. The system proved to be robust even when faced with data from different populations and clinical settings. By providing a Pareto-optimal balance between AI autonomy, human expert effort, and diagnostic accuracy, the MPD$^2$-Router offers a practical path toward integrating AI into real-world ophthalmic workflows.

Comments (0)

No comments yet

Be the first to share your thoughts!