Evaluating cognitive biases in AI-assisted mammography interpretation: a simulation reader study of explainable AI across radiologist experience levels.
Authors
Affiliations (5)
Affiliations (5)
- Breast Imaging Division, Radiology Department, IEO European Institute of Oncology IRCCS, Milan, Italy. [email protected].
- Breast Imaging Division, Radiology Department, IEO European Institute of Oncology IRCCS, Milan, Italy.
- Department of Psychology, Educational Science and Human Movement, University of Palermo, Palermo, Italy.
- Applied Research Division for Cognitive and Psychological Science, European Institute of Oncology, IRCCS, Milan, Italy.
- Department of Oncology and Hemato-Oncology, University of Milan, Milan, Italy.
Abstract
To evaluate the impact of automation and anchoring bias in artificial intelligence (AI)-assisted mammography interpretation and to assess whether saliency-based explainable AI (XAI) mitigates these biases across radiologists of varying experience. In this monocentric, fully crossed simulation reader study conducted between March and June 2024, six breast radiologists stratified by experience independently reviewed 200 mammograms under three sequential conditions: unassisted, AI-assisted, and AI-assisted with saliency-based XAI heatmaps. To quantify susceptibility to misleading AI advice under controlled discordance conditions, BI-RADS-like AI recommendations were deliberately perturbed by one category in 30% of examinations, whereas the remaining 70% retained the native AI output. Bias outcomes were analyzed using generalized linear mixed-effects models accounting for reader- and case-level clustering. In the AI-assisted condition without explanations, automation bias occurred in 65/180 (36.1%) and anchoring bias in 61/180 (33.9%) of manipulated cases. With XAI, these rates decreased to 32/180 (17.8%) and 31/180 (17.2%), respectively. In mixed-effects models, XAI was associated with lower odds of automation bias (aOR 0.56, 95% CI 0.44-0.71; p < 0.001) and anchoring-related revision bias (aOR 0.61, 95% CI 0.48-0.78; p < 0.001). On the non-manipulated subset, diagnostic accuracy improved from 724/840 (86.2%) in the unaided phase to 757/840 (90.1%) in the AI + XAI phase. Automation and anchoring bias affected AI-assisted mammography interpretation, particularly among less experienced radiologists. Saliency-based explainable AI reduced, but did not eliminate, these effects. Question AI assistance can systematically influence BI-RADS decisions in mammography, particularly among less experienced radiologists, through automation and anchoring biases. Findings Saliency-based explainable AI (XAI) substantially reduces biased decisions while modestly improving overall diagnostic accuracy compared with standard AI support alone. Clinical relevance Embedding XAI and targeted training into AI-assisted mammography workflows may enhance patient safety and support safer clinical integration of mammography AI tools.