Year: 2025
Type: article
Abstract: Artificial intelligence is advancing quickly in radiology, yet average accuracy and black-box scores can hide per-case risk. This viewpoint argues that confidence is a safety feature, not a cosmetic add-on. Actionable confidence rests on three elements—calibration so a score means what it says, uncertainty to signal when a case may fail, and deferral so the system can hand a study back cleanly. We set a clinician-friendly display principle that uses simple standardized cues—text or color where appropriate—shown sparingly and only at validated extremes, while mid-confidence cases stay quiet with detail available on demand. Applied across screening and triage, detection and segmentation, quantification and longitudinal change, opportunistic alerts, and report phrasing, this approach improves prioritization and workload without flooding readers with numbers. Equally important, it bridges developers and radiologists by aligning on the clinical role and operating points, verifying calibration on local data, and monitoring a light post-deployment set that tracks what the AI handles, residual risk in acted-upon cases, and dataset drift. Used this way, confidence clarifies when the model knows, when it does not, and why, turning scores into trusted decisions that accelerate care while guarding against automation bias.
Author: Nalan Karunanayake
Institution: Memorial Sloan Kettering Cancer Center
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $2,900
OpenAlex ID: W4415319924