Radiology

Hard–easy effect and time dynamics drive overconfidence in radiologic reads

In a 2806-decision study of 28 radiologists and 50 nonradiologists reviewing bone radiographs, radiologists had higher accuracy (73.6%) and confidence than nonradiologists. Overall overconfidence affected 14.73% of decisions and was larger for difficult cases, showing a clear hard–easy effect with >40% overconfidence on hard images. Fast decisions were accurate and well calibrated, whereas prolonged deliberation predicted more errors and increased overconfidence. Lesion-present calls had lower confidence and more frequent deferral to seniors, suggesting calibrated uncertainty should prompt consultation. 1

Saliency-based XAI halves automation bias in AI-assisted mammography

Six breast radiologists read 200 mammograms in unaided, AI-assisted, and AI+saliency XAI conditions with deliberate AI perturbations in 30% of cases. Automation bias fell from 65/180 (36.1%) without XAI to 32/180 (17.8%) with XAI, and anchoring reversions showed similar reductions. In non-manipulated cases, diagnostic accuracy improved from 86.2% unaided to 90.1% with AI+XAI. XAI reduced but did not eliminate biased decisions, especially among less-experienced readers, implying training remains necessary. 2

Patients accept AI if transparency, oversight, and humane care are preserved

Seven focus groups (34 patients, mean age 49.1 years) expressed conditional openness to AI in radiology contingent on evidence of safety, added value, and faster results. Patients require transparency about AI use, binding quality assurance, and explicit disclosure to consent to AI-assisted diagnostics. They expect independent physician judgment, clear explanation of AI limits, and preserved empathetic clinician–patient interactions. Failure to preserve physician responsibility risks loss of trust and perceived dehumanization of care. 3