Characterize the clinician-response distribution for safe query control

Characterize the response distribution for clinician feedback in clinician-in-the-loop few-shot medical image analysis, including disagreement, nonresponse, latency, and interface effects, so that query policies do not rely on misspecified response models.

Background

The decision-theoretic formulation assigns value to clinician queries using a response model that predicts how feedback will affect subsequent risk and decisions. The paper notes that clinician responses are fallible and can vary because of disagreement, nonresponse, latency, and interface conditions. Accurate modeling of these effects is necessary for estimating the value of information and preventing confidently low-value queries.

References

The response distribution $p(z\mid s,a)$ is another unresolved source of error.

— From Few-Shot Segmentation to Clinician-in-the-Loop Medical Image Analysis  (2609.10001 - Zhu, 9 Sep 2026) in Section 9, “Limitations, Governance, and Scope”

The simulated clicks always target the deepest error of the current prediction, whereas an annotator may click elsewhere. How the update responds to clicks that do not target the deepest error remains to be measured, and a reader study with clicks chosen by clinicians is therefore the natural next step.

— LeCor: Learning to Be Corrected by Meta-Learned Test-Time Training for Interactive 3D Lung-Tumour Segmentation  (2609.09477 - Luo et al., 8 Sep 2026) in Section Discussion