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

2026-05-08 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research · Depth: Expert, quick

Summary

The MPD$^2$-Router is a mask-aware multi-expert deferral framework designed to improve glaucoma screening safety by routing difficult or uncertain cases to human experts. This system recasts ophthalmic triage as a constrained human-AI routing problem, determining whether to defer a case and to which available expert. It integrates a dual-head deferral/allocation policy with mask-aware Gumbel-sigmoid gating to enforce per-sample availability, while fusing uncertainty, morphology, image-quality, and out-of-distribution signals. The training employs an asymmetric cost-sensitive objective with an augmented-Lagrangian deferral budget, a group-specific distribution prior, and a rank-majorization JS regularizer to prevent expert collapse without forcing uniform allocation. Evaluated across three cross-national glaucoma cohorts (REFUGE, CHAKSU, ORIGA) using a frozen REFUGE-trained backbone, MPD$^2$-Router significantly reduces clinical cost and enhances Matthews Correlation Coefficient (MCC) compared to AI-only approaches at a moderate deferral rate, demonstrating Pareto-optimality in F1-MCC-cost and robustness under cross-domain shift.

Key takeaway

For Computer Vision Engineers developing medical diagnostic AI, consider implementing a deferral router like MPD$^2$-Router to enhance safety and efficiency. Your system should account for expert availability, workload balance, and asymmetric diagnostic costs to achieve Pareto-optimal performance and robustness across diverse patient cohorts. This approach can significantly lower clinical costs and improve diagnostic accuracy by intelligently routing challenging cases to human oversight.

Key insights

MPD$^2$-Router optimizes glaucoma screening by routing complex cases to available human experts, balancing cost and diagnostic accuracy.

Principles

• Integrate multiple signals for robust case difficulty assessment.
• Enforce expert availability for practical deferral systems.
• Balance expert workload to prevent collapse.

Method

The MPD$^2$-Router uses a dual-head deferral/allocation policy with mask-aware Gumbel-sigmoid gating, fusing uncertainty, morphology, image-quality, and OOD signals. Training involves an asymmetric cost-sensitive objective, augmented-Lagrangian budget, group-specific prior, and rank-majorization JS regularizer.

In practice

• Apply mask-aware gating for resource-constrained routing.
• Utilize asymmetric cost functions for medical diagnostics.
• Combine multiple uncertainty signals for robust deferral.

Topics

• Glaucoma Screening
• Learning-to-Defer
• Multi-expert AI
• MPD$^2$-Router
• Ophthalmic Triage

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Bridging the Rural Healthcare Gap: A Cascaded Edge-Cloud Architecture for Automated Retinal Screening — An edge-cloud cascade reduces cloud usage for DR screening while maintaining high diagnostic performance.
• EyeMVP: OCT-Informed Fundus Representation Learning via Paired CFP--OCT Pretraining — EyeMVP uses paired CFP-OCT pretraining to enrich CFP representations with OCT depth information for improved retinal analysis.
• QG-MIL: A Gated Transformer Aggregator for Domain-Agnostic Multiple Instance Learning in Medical Imaging — QG-MIL stabilizes attention in MIL aggregators for medical imaging, improving prediction stability and cross-domain performance.
• Towards Real-Time Autonomous Navigation: Transformer-Based Catheter Tip Tracking in Fluoroscopy — A multi-threaded deep learning pipeline enables robust, real-time catheter tip tracking in fluoroscopy for autonomous navigation.
• DynFS-MoE: Dynamic Functional-Structural Mixture-of-Experts for Post-Traumatic Epilepsy Diagnosis — DynFS-MoE dynamically integrates multimodal MRI data via expert routing for interpretable, early post-traumatic epilepsy diagnosis.
• Validating a Deep Learning Algorithm to Identify Patients with Glaucoma using Systemic Electronic Health Records — An EHR-only deep learning model can effectively pre-screen for glaucoma without specialized imaging.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.