OPERA framework enables universal biomedical image analysis without retraining

Biomedical image analysis models often fail in real-world deployment due to distribution shifts in scanners, protocols, and patient populations. Traditional solutions require costly domain-specific fine-tuning, which is impractical when labels are scarce or privacy restricts data sharing. OPERA (Offline Policy-guided Expert Routing and Adaptation), presented by Zihan Li et al. and accepted at ACM MM 2026, tackles this by treating expert weight assignment as an offline policy learning problem. It learns a routing policy from a small validation set without gradient updates to individual expert agents, then deploys with test-time adaptation. The framework includes expert profiling (offline selection policies), confidence calibration via temperature adjustment, and distribution-aware adaptation that dynamically adjusts class weights using unlabeled test data batch statistics. Instance-level routing assigns each sample to the best expert based on inter-model agreement and predictive entropy.

OPERA was tested on 9 diverse datasets covering fundus photography, chest X-ray, CT, MRI, and multimodal diagnostic benchmarks. It outperformed 30+ baseline models in classification, segmentation, and multimodal settings. Key results show consistent improvements in both prediction accuracy and calibration quality, demonstrating that offline policy-guided coordination of heterogeneous specialists offers a practical path to deployable biomedical AI without retraining. The framework works with existing pre-trained models and requires no changes to underlying architectures. Code is available on GitHub. This approach could significantly reduce the engineering overhead of adapting medical AI systems to new clinical environments.