Calibrated Triage, Not Autonomy: Confidence Estimation for Medical Vision-Language Models

2026-06-14

AI Summary

This work addresses the tendency of medical vision-language models to generate plausible yet incorrect answers due to reliance on linguistic priors, coupled with the absence of reliable confidence estimation for deciding when to answer autonomously or defer to clinicians. The authors reformulate medical visual question answering as a bounded selective prediction task, wherein the model responds only when its confidence exceeds a threshold, otherwise deferring to human experts. Introducing high-confidence error rate as a key evaluation metric, they expose the limitations of conventional metrics in real-world deployment scenarios and systematically assess seven confidence estimation methods across five open-source vision-language models and three medical VQA datasets—all models trained solely on natural images and calibrated via out-of-domain temperature scaling. The best-performing method reduces high-confidence error rates to 1–4% (from baselines of 41–45%), enabling safe automation of approximately one-third of radiology cases under a 20% error tolerance, whereas pathology tasks remain largely non-automatable, highlighting that practical deployability is jointly constrained by model capability and confidence calibration.

Abstract

A vision-language model can answer a question about a medical image fluently and confidently while barely using the image, leaning instead on language priors. In medicine this is the failure that matters most, because the answer looks trustworthy and is not, and the only protection is a confidence score reliable enough to tell the system when to abstain. We ask a deployment question rather than an accuracy one: how much imaging work a model can safely handle alone, and which confidence signal makes that possible. We evaluate seven confidence estimators across five open-weight LVLMs and three medical visual-question-answering datasets spanning broad clinical imaging, radiology, and pathology, with every probe trained only on natural images and applied without adaptation. Recast as bounded selective prediction (automate a case only when confidence clears a threshold, defer the rest), the comparison is cautionary. The standard metrics are poor guides: discrimination barely separates the methods, and the weak calibration of a cheap self-report is cheaply removed by off-domain temperature scaling without changing deployable yield. What distinguishes a usable estimator is the high-confidence region a clinician acts on: the weakest baselines are confidently wrong on 41 to 45 percent of their errors against 1 to 4 percent for the best probe, and no estimator is reliably best across domains or models. Safe handoff is governed at two levels: base-model competence sets a ceiling, so a well-calibrated score recovers roughly a third of radiology cases at a 20 percent error tolerance but almost none of pathology; the confidence layer then decides how much of that ceiling is reachable. The usable role today is calibrated triage, not autonomy: automate the cases a calibrated score marks safe, route the rest to a clinician. We release all outputs, correctness judgments, and confidence scores, with code.

Problem

• confidence estimation
• medical vision-language models
• calibrated triage
• selective prediction
• model reliability

Innovation

• confidence estimation
• medical vision-language models
• calibrated triage
• selective prediction
• trustworthy AI