🤖 AI

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

The paper introduces MPD $^2$ -Router, a novel mask-aware multi-expert deferral framework that optimizes glaucoma screening by dynamically routing uncertain cases to available human experts based on morphology, uncertainty, and cost constraints, thereby achieving Pareto-optimal performance in clinical cost, accuracy, and balanced workload distribution across diverse cohorts.

Original authors: Wenxin Zhan

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wenxin Zhan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are running a busy, high-stakes triage center for eye health. You have a very fast, very smart computer (the AI) that looks at eye photos and tries to spot glaucoma. But you also know that the computer isn't perfect; sometimes it gets confused, especially if the photo is blurry or the eye looks unusual.

In the past, the standard approach was simple: "Let the computer decide everything." If the computer was unsure, it would just guess anyway, which could be dangerous. Or, the old way of "deferring" (asking a human for help) was like a blunt instrument: "If the computer is unsure, send everything to a human." This is inefficient because it clogs up the system, and it treats all human doctors as if they were identical clones, which they aren't.

This paper introduces MPD2-Router, a new "traffic cop" for your eye clinic. Here is how it works, using simple analogies:

1. The Problem: The "One-Size-Fits-All" Mistake

Think of the old AI systems as a single-lane highway. Every car (patient) goes through the same gate. If the gate is broken (the AI is unsure), the traffic jams up, or cars crash because the gate forced them through anyway.
Also, the old systems treated all human doctors like interchangeable batteries. They didn't realize that Dr. Smith is great at spotting early signs but bad at blurry photos, while Dr. Jones is the opposite. Sending a blurry photo to Dr. Smith is a waste of time; sending it to Dr. Jones is smart.

2. The Solution: The "Smart Traffic Cop" (MPD2-Router)

MPD2-Router is a new system that sits between the AI and the human doctors. It doesn't just ask, "Is the AI unsure?" It asks a much more detailed set of questions:

Is the photo clear? (Image Quality)
Does the eye look weird or unusual? (Out-of-Distribution signals)
Does the eye structure look suspicious? (Morphology)
Who is actually available right now? (Availability Mask)

The "Mask" Analogy:
Imagine you have a panel of 12 different doctors. But for any specific patient, maybe only 3 of them are on duty, and the other 9 are at lunch or on vacation. The "Mask" is a digital list that says, "For this patient, you can only ask Doctors A, B, and C." The system strictly forbids sending the patient to Doctor D, even if the computer thinks Doctor D is the best.

3. How It Decides (The Two-Headed Brain)

The system has two distinct decisions to make, like a two-headed coin:

Head 1: "Should we stop?" It decides if the AI should keep the case or hand it over to a human. It doesn't just guess; it weighs the risk of a mistake against the cost of asking a human.
Head 2: "Who should we ask?" If it decides to hand it over, it looks at the "Mask" (who is available) and picks the best available doctor for that specific type of problem. It doesn't just pick the first one; it matches the patient's specific needs to the doctor's specific strengths.

4. Keeping the System Balanced (The "Anti-Collapse" Rules)

A common problem in these systems is "Expert Collapse." Imagine if the system gets so smart that it realizes Dr. Jones is the best at everything. It would send every single difficult case to Dr. Jones. Dr. Jones would get overwhelmed, while the other 11 doctors sit idle.
MPD2-Router has special rules (regularizers) to prevent this:

The "Group" Rule: It ensures that if a group of patients has similar eye structures, the cases are spread out among the available doctors, not dumped on one person.
The "Rank" Rule: It makes sure the system doesn't just pick the "top" doctor for every single case. It forces the system to use the whole team, keeping the workload balanced.

5. The Results: Safer and Smarter

The authors tested this system on three different groups of patients from different countries (different "universes" of eye data).

The AI alone made many mistakes, especially on the tricky, unusual cases.
Old deferral systems either sent too many people to humans (wasting time) or sent them to the wrong humans.
MPD2-Router found the "sweet spot." It kept the AI in charge for the easy cases (saving time) but perfectly identified the hard cases and routed them to the right available human doctor.

The Bottom Line:
The paper claims that MPD2-Router is like a smart, fair, and safety-conscious manager. It knows when to trust the computer, when to call a human, and exactly which human to call based on who is available and who is best suited for the job. This leads to fewer missed diagnoses (safer) and less wasted time for the doctors (more efficient), without overloading any single specialist.

Technical Summary: MPD2-Router

Problem Statement

The paper addresses the limitations of standard "one-shot" AI predictors in ophthalmic glaucoma screening. Current systems typically output a single label without considering the sequential, uncertain nature of clinical decision-making. The authors identify several critical gaps in existing Learning-to-Defer (L2D) formulations when applied to real-world clinical workflows:

Expert Heterogeneity: Human experts (ophthalmologists, trainees, graders) are not interchangeable oracles; they vary in experience, subspecialty knowledge, and decision thresholds.
Availability Constraints: In practice, not all experts are available for every case, yet standard L2D often assumes a static, fully available expert set.
Asymmetric Harm: Diagnostic errors have asymmetric costs (e.g., missing glaucoma is more harmful than a false referral), which standard formulations often overlook.
Distribution Shift and Morphology: AI reliability degrades under out-of-distribution (OOD) conditions, low image quality, and specific morphological difficulties (e.g., cup-to-disc ratio variations), yet standard models often rely solely on confidence scores.
Expert Collapse: Existing multi-expert routing methods often fail to distribute workload evenly, causing "expert collapse" where difficult cases are routed to a single dominant expert, overloading them.

The core problem is reframed as a constrained human–AI routing problem: determining not only whether to defer a case to a human but also which available expert is most appropriate, while managing workload, cost, and clinical safety.

Methodology: MPD2-Router

The authors propose MPD2-Router (Mask-aware Multi-expert Prior-regularized Dual-head Deferral Router), a framework that couples a dual-head policy with mask-aware gating and specialized regularization.

1. Router Representation

The router ingests a fused representation ( $h_i$ ) derived from three complementary sources:

Frozen AI Logits: Outputs from a pre-trained, frozen SwinV2 classifier.
Risk Signals: A vector combining Virtual-logit Matching (VIM) scores, image-quality risk, and maximum-probability uncertainty.
Morphological Features: Structural features derived from optic-disc and optic-cup segmentation (vertical and area cup-to-disc ratios), including a structural risk estimate and a measure of AI-structure disagreement.

2. Dual-Head Policy

The policy $\pi_\theta$ is factorized into two coupled decisions:

Deferral Head: Estimates the soft deferral mass $d_i$ (probability of deferring to a human). It enforces autonomous prediction if no experts are feasible for a given sample.
Expert Allocation Head: Determines to whom to defer. This head uses a mask-aware Gumbel–sigmoid gating mechanism:
- Stage I (Support Selection): Uses an extended-real masking map to strictly enforce feasibility. Infeasible experts are masked to $-\infty$ before the Gumbel-softmax relaxation, ensuring the probability of selecting an unavailable expert is exactly zero.
- Stage II (Allocation): Allocates probability mass only over the set of feasible experts selected in Stage I, normalized via a masked softmax.
- Repair Mechanism: If the stochastic support selection results in an empty set, the policy falls back to the full feasible mask to ensure a valid action.

3. Regularization and Objectives

To prevent expert collapse and ensure balanced workload, the training objective includes:

Group-Specific Distribution Prior (GSDP): A hierarchical empirical-Bayes-style prior that shrinks group-level allocation distributions toward reliability-weighted estimates derived from global cohort statistics. This prevents specific groups of cases from over-concentrating on a single expert.
Rank-Majorization Jensen–Shannon (JS) Regularizer: Penalizes the sorted rank profile of the allocation distribution. It activates only when the router assigns excessive cumulative mass to the top-ranked experts relative to a geometric reference, discouraging sample-level concentration without forcing uniformity.
Asymmetric Cost-Sensitive Objective: Minimizes a clinical-operational risk function that accounts for false-negative/positive costs ( $c_{fn}, c_{fp}$ ) and operational expert costs ( $\kappa_j$ ).
Augmented-Lagrangian Deferral Control: Enforces a soft deferral budget ( $\bar{d} \le \rho_{def}$ ) using an augmented-Lagrangian multiplier, allowing the model to learn a controlled deferral rate rather than an uncontrolled byproduct.

Key Contributions

Formulation: Recasts glaucoma triage as a constrained routing problem that explicitly models per-sample expert availability, heterogeneous expert behavior, and asymmetric diagnostic harm.
Architecture: Introduces a dual-head router with mask-aware Gumbel–sigmoid gating that strictly enforces feasibility constraints during both training and inference, separating the "whether to defer" decision from the "whom to defer to" decision.
Regularization: Proposes a novel combination of GSDP and Rank-Majorization JS regularizers to mitigate expert collapse (concentration of deferred cases on few experts) while preserving cost-aware allocation.
Training Strategy: Utilizes an asymmetric cost-sensitive objective with an augmented-Lagrangian controller to yield clinically aligned decisions under a deferral budget.

Experimental Results

The method was evaluated on three cross-national glaucoma cohorts: REFUGE, CHAKSU, and ORIGA. A frozen AI backbone trained on REFUGE was used to simulate cross-domain deployment shifts.

Performance: MPD2-Router achieved the strongest overall performance, improving accuracy from 0.755 (AI-only) to 0.960, F1 from 0.530 to 0.868, and MCC from 0.482 to 0.844.
Cost Efficiency: It significantly reduced clinical cost (from 0.372 to 0.068) and total cost compared to AI-only and other baselines (e.g., Verma-OvA, Hemmer-MoE).
Pareto Optimality: The method occupies a Pareto-favorable region in the F1–MCC–cost trade-off space, achieving high predictive utility without incurring the excessive deferral costs seen in other methods.
Robustness: Under distribution shift (specifically on the ORIGA dataset where AI-only accuracy dropped to 0.491), MPD2-Router recovered performance to 0.926 accuracy and 0.820 MCC while maintaining the lowest total cost.
Expert Utilization: The method demonstrated balanced expert utilization, achieving the lowest Top-1 share (24.7%) and Top-2 share (44.0%) among compared methods, indicating it avoids routing collapse onto a single dominant expert.
Risk Stratification: The router exhibited slower cost growth under high structural and model-reliability risk compared to baselines, remaining substantially below AI-only performance in the highest-risk strata.

Significance and Claims

The paper claims that MPD2-Router provides a clinically grounded framework for analyzing when autonomous AI is appropriate and how scarce expert attention should be allocated safely under distribution shift.

Safety and Selectivity: The system is selective rather than indiscriminate, routing cases based on a fusion of uncertainty, morphology, image quality, and OOD signals, rather than just confidence.
Practical Deployment: By enforcing per-sample availability and preventing expert collapse, the method addresses key deployment failures of standard L2D, such as overloading specific specialists or ignoring availability constraints.
Limitations: The authors modestly note that the evaluation relies on retrospective cross-cohort datasets with heterogeneous and partially observed expert labels (using semi-synthetic labels for REFUGE and pseudo-labels for ORIGA). They state that prospective clinical validation and regulatory review are necessary before actual deployment, but the cross-cohort results suggest the method remains effective under sparse and heterogeneous supervision.

The work positions structured human–AI routing as a practical direction for safer ophthalmic AI deployment, moving beyond simple classification to instance-wise, cost-aware, and availability-aware decision support.