Evo-RAD: Navigating Rare Retinal Disease Diagnosis via Self-Evolving Agentic Retrieval

Large-scale foundation models often struggle with rare retinal diseases due to data scarcity and the tendency to confuse rare conditions with visually similar common ones. Existing retrieval-augmented diagnosis (RAD) methods typically rely on static nearest-neighbor searches, which are prone to the 'hubness problem'—where common, irrelevant samples are repeatedly retrieved. This paper asks whether a dynamic, agentic approach to evidence retrieval can improve diagnostic accuracy for rare conditions.

Evo-RAD treats the retrieval process as a Markov Decision Process (MDP). A graph-based agent observes the current set of retrieved reference images and performs sequential operations: deleting discordant evidence (DELETE), adding pathologically consistent samples from a candidate buffer (INSERT), or stopping the process (TERMINATE). The agent is trained using Group Relative Policy Optimization (GRPO) with a custom 'homogeneity-aware' reward function that encourages the final reference set to be both diagnostically pure and semantically consistent with the query image.

Evo-RAD significantly outperforms existing retinal foundation models and standard retrieval methods. On the Rare-20 benchmark, it achieved a 46.28% accuracy, representing a substantial improvement over zero-shot foundation models (+21.04%) and state-of-the-art static retrieval or parameter-efficient fine-tuning (PEFT) methods (+3.56%). The authors demonstrate that the agent's ability to iteratively refine the evidence set—rather than relying on a single-pass retrieval—is the primary driver of these gains.

This work shifts the paradigm of medical image retrieval from passive matching to active, agentic reasoning. By mimicking the clinical process of hypothesis refinement, Evo-RAD provides a robust, parameter-efficient way to adapt foundation models for rare disease diagnosis without requiring extensive retraining on imbalanced datasets.