OPERA: Offline Policy-guided Expert Routing and Adaptation for Universal Biomedical Image Analysis

Analysis

Why This Paper Matters

Biomedical image analysis faces a critical deployment bottleneck: models trained on one dataset often fail when applied to new scanners, protocols, or patient populations. Traditional solutions require repeated domain-specific fine-tuning, which is costly, label-hungry, and often infeasible due to privacy constraints. OPERA addresses this by introducing a multi-agent ensemble framework that learns to route test samples to the most suitable expert without any gradient updates to the expert models. This offline policy-guided approach is a significant step toward practical, deployable biomedical AI that can handle distribution shifts without retraining.

The paper is particularly relevant for AI practitioners working on real-world clinical systems where data heterogeneity is the norm. By treating expert assignment as an offline policy learning problem, OPERA sidesteps the computational and data demands of online adaptation methods. The framework's ability to work with unlabeled test data and a small validation set makes it attractive for privacy-sensitive and label-scarce environments.

Technical Contributions

• Offline Policy-guided Expert Routing: Learns a routing policy from a small validation set without updating expert agents, enabling efficient deployment.
• Expert Profiling Module: Selects policies offline to allocate expertise based on validation performance.
• Confidence Calibration: Adjusts temperature scaling for each expert to produce more reliable probabilistic outputs.
• Distribution-aware Adaptation: Dynamically adjusts class weights at the batch level using statistics from unlabeled test data.
• Instance-level Routing: Assigns each sample to the most suitable expert using inter-model agreement and predictive entropy.

Results

OPERA was evaluated on 9 diverse datasets covering fundus photography, chest X-ray, CT, MRI, and multimodal diagnostic benchmarks. It was compared against over 30 baselines across classification, segmentation, and multimodal settings. The framework consistently improved both performance (e.g., accuracy, Dice score) and calibration quality (e.g., expected calibration error). The results demonstrate that offline policy-guided coordination of expert agents is a viable alternative to retraining for handling distribution shifts.

Significance

OPERA offers a practical blueprint for building robust biomedical AI systems that can be deployed across heterogeneous clinical environments without costly retraining. By decoupling policy learning from expert model updates, it enables the reuse of pre-trained specialist models in new settings. This approach could accelerate the adoption of AI in clinical workflows where data diversity and privacy constraints are major hurdles. The framework's reliance on unlabeled test data for adaptation further enhances its real-world applicability.