Related Experiment Video
Updated: Jul 31, 2026
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
Automation Bias in Action: Eye Tracking of Humans Reading Screening Mammograms with and without AI Prompts
Adnan G Taib1, George J W Partridge1, Peter Phillips2
1Translational Medical Sciences, School of Medicine, University of Nottingham, Clinical Sciences Building, City Hospital Campus, Nottingham City Hospital, Hucknall Rd, Nottingham NG5 1PB, UK.
Radiology
|July 14, 2026
Summary
Incorrect artificial intelligence (AI) suggestions in mammography reduce diagnostic accuracy, particularly with false-negative findings. AI prompts increase reading time and alter visual search patterns, necessitating careful AI calibration.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Automation bias from incorrect artificial intelligence (AI) suggestions is a concern in medical image interpretation.
- The impact of AI errors on mammography interpretation accuracy and reader behavior is under-researched.
Purpose of the Study:
- To evaluate how incorrect AI suggestions affect diagnostic accuracy, reading times, and visual search behavior in mammography interpretation.
- To assess the influence of true-positive, false-negative, false-positive, and true-negative AI suggestions on reader performance.
Main Methods:
- A retrospective multireader paired study involving 10 mammography readers.
- Readers interpreted screening mammograms with and without AI decision support.
- Eye-tracking recorded visual search behavior; statistical tests compared outcomes.
Main Results:
- False-negative AI suggestions significantly decreased reader sensitivity (39% with AI vs. 71% without).
- False-positive AI suggestions increased specificity (39% vs. 21%) but led to shorter fixation durations.
- Increased AI prompts correlated with longer reading times and reduced fixations on missed cancers.
Conclusions:
- Incorrect AI suggestions significantly impact reader accuracy and visual search behavior in mammography.
- False-negative AI suggestions pose the greatest risk, highlighting the need for AI threshold calibration.
- AI tools require careful implementation to mitigate automation bias and maintain diagnostic performance.
