Retinal screening stops sending every case to review: RetinaGuard accepts about 37% of cases at a 5% unsafe level using evidence-gathering agents and conformal risk control

An automated retinal screening decision can now carry a certified upper bound on its unsafe rate, rather than only ranking cases by confidence and sending the least certain ones to an ophthalmologist. The prior approach ranked cases by confidence or calibrated uncertainty and referred the least certain; recent agentic clinical systems call tools but are rarely compared with confidence-based referral at the same referral cost, and their decisions carry no statistical guarantee. On the 640 images of the official RFMiD test split, all methods are compared at the same six referral budgets of 5, 10, 15, 20, 30 and 40%, with an equivalent certified safety level from the same conformal calibration.

The system pools each fundus photograph's evidence into a probability that the automated decision is unsafe, then decides whether to accept, gather more evidence or refer. Here an agent is a component that decides where a case goes next or whether computation is worth spending, covering quality, diagnostic, uncertainty, verification, evidence, decision and reporting stages. The diagnostic stage fine-tunes RETFound for multi-label probabilities, the uncertainty stage runs 20 stochastic dropout passes, and the verification stage uses an independently trained ConvNeXt-B whose prediction never replaces the primary one and only feeds disagreement into the risk model.

The verifier and explanation tools are invoked only when their evidence is expected to change the decision enough to justify the measured latency, and the acceptance threshold is certified on a held-out calibration set. Wrapping tools around a classifier helps only if the evidence they produce improves the accept-or-refer decision, and whether it does depends on how much the primary model's own uncertainty already reveals about its errors. The gate calls the verifier for 24.8% of test cases when disagreement is informative and for 7.7% in the stronger-model setting where it is not, cutting computation by 30% and 29% respectively.

The conclusions rest on the 640 images of the official RFMiD test split; all development data come from this one public dataset, and external evaluation uses ODIR-5K labels derived from diagnostic keywords without retraining. The intended readers are researchers and engineers designing screening triage and selective prediction; the source proposes next steps of measuring the dependence on model strength systematically across primary models and verifiers, validating the quality agent against clinician-graded gradability, introducing clinically weighted losses and conformal methods that adapt to distribution shift, and evaluating on larger multi-site cohorts.

The conformal guarantee is marginal, assumes exchangeability and does not hold under population shift; the two primary models differ only in fine-tuning schedule, so the dependence on model strength rests on two points rather than a systematic sweep; the quality agent was validated with synthetic corruptions rather than graded gradability labels; the unsafe-decision definition weights all missed labels equally, although cost-weighted losses fit the same framework; and computational cost comes from measured per-tool latencies rather than an end-to-end deployment.