Automation Bias in AI-Assisted Detection of Cerebral Aneurysms on Time-of-Flight MR-Angiography

Su Hwan Kim, Severin Schramm, Evamaria Olga Riedel, Lena Schmitzer, Enrike Rosenkranz, Olivia Kertels, Jannis Bodden, Karolin J. Paprottka, Dominik Sepp, Martin Renz, Jan S. Kirschke, Thomas Baum, Christian Maegerlein, Tobias Boeckh‐Behrens, Claus Zimmer, Benedikt Wiestler, Dennis M. HedderichmedRxiv 2024 · DOI

Abstract

AI systems have the potential to support in detecting cerebral aneurysms. Yet, the role of automation bias (inclination of humans to overly rely on automated decision-making systems) in AI-assisted cerebral aneurysm detection remains unclear. Purpose: To determine how automation bias can affect radiologists with varying experience levels when reading time-of-flight magnetic resonance angiography (TOF-MRA) studies with the assistance of an AI system for cerebral aneurysm detection. Methods: In this prospective experiment, nine radiologists with varying levels of experience evaluated twenty TOF-MRA exams for the presence of anterior circulation aneurysms, with each arterial segment rated on a 4-point Likert scale, and provided follow-up recommendations. Every case was evaluated twice (with or without assistance by the AI software © mdbrain), with a washout-period of at least four weeks between the two sessions. Ten out of twenty cases included at least one false-positive AI finding. Aneurysm ratings, follow-up recommendations, and reading times were assessed using the Wilcoxon signed-rank test. A thematic analysis was performed to summarize reader feedback and observations. Results: False-positive AI results led to significantly higher suspicion of aneurysm findings (p = 0.01). Inexperienced readers further recommended significantly more aggressive follow-up examinations when presented with false-positive AI findings (p = 0.005). Reading times were significantly shorter with AI assistance in inexperienced (164.1 vs 228.2 seconds; p < 0.001), moderately experienced (126.2 vs 156.5 seconds; p < 0.009), and very experienced (117.9 vs 153.5 seconds; p < 0.001) readers alike. Conclusion: Our results demonstrate susceptibility of radiology readers to automation bias in detecting cerebral aneurysms in TOF-MRA studies when encountering false-positive AI findings. In inexperienced readers, this behavior further translated into more aggressive follow-up recommendations. AI assistance resulted in significantly shorter reading times across experience levels. While AI systems for cerebral aneurysm detection can provide benefits, challenges in human-AI interaction need to be mitigated to ensure safe and effective adoption.

Introduction

Cerebral aneurysms are localized dilations of brain arteries, affecting approximately 2% of the population. While often asymptomatic, their rupture can lead to life-threatening subarachnoid hemorrhages. Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) is a standard non-invasive tool for detection, but the process is cognitively demanding and prone to human error.

Artificial Intelligence (AI) and Computer-Assisted Diagnosis (CAD) systems are increasingly integrated into clinical workflows to improve sensitivity. However, these systems often prioritize high sensitivity to avoid missing lesions, leading to a high rate of false positives. This study investigates automation bias—the tendency of human operators to over-rely on automated suggestions—and how it influences radiologists of varying experience levels in the context of aneurysm detection.

Methodology

The researchers conducted a prospective crossover study involving nine radiologists categorized into three experience tiers:

  • Inexperienced: Residents with less than one year of neuroradiology experience.
  • Moderately Experienced: Board-certified radiologists.
  • Very Experienced: Board-certified neuroradiologists.

Participants evaluated 20 TOF-MRA cases twice: once with AI assistance (using mdbrain software) and once without. A four-week washout period separated the sessions to minimize recall bias. The AI acted as a "first reader," providing annotations for the radiologists to review before their own systematic assessment.

Dataset and Ground Truth

The dataset comprised 20 3D TOF-MRA studies from an outpatient practice.

  • 10 cases contained at least one true-positive aneurysm.
  • 10 cases contained at least one false-positive AI finding (e.g., vascular loops or infundibula).

Ground truth was established by consensus between two senior neuroradiologists (16 and 9 years of experience). Readers rated each arterial segment on a 4-point Likert scale and recommended follow-up actions ranging from no action to invasive Digital Subtraction Angiography (DSA).

Key Results

The study revealed a significant impact of AI assistance on both diagnostic behavior and efficiency.

Sensitivity and Specificity

Overall sensitivity for aneurysm detection improved from 0.88 to 0.97 with AI assistance. However, this gain came at a cost. False-positive AI findings significantly increased the suspicion of aneurysms among readers ($p = 0.01$). For inexperienced readers, specificity dropped from 0.81 to 0.66, as they were more likely to accept false AI flags as real findings.

Follow-up Recommendations

Automation bias translated into clinical decision-making. Inexperienced readers recommended significantly more aggressive follow-up procedures (e.g., MRI or DSA) when presented with false-positive AI findings ($p = 0.005$). In contrast, more experienced readers were better able to dismiss false positives, suggesting that clinical experience serves as a protective factor against automation bias.

Efficiency

AI assistance significantly reduced reading times across all experience levels. Mean reading times dropped from 179.4 seconds to 136.0 seconds. Even the most experienced neuroradiologists saw a reduction from 153.5s to 117.9s.

Discussion

The findings highlight several critical aspects of human-AI interaction in radiology.

The Experience Buffer

The susceptibility to automation bias was inversely proportional to experience. While junior residents often deferred to the AI's judgment, experts maintained a higher level of diagnostic skepticism. This suggests that AI tools may be most dangerous in the hands of those they are intended to help the most—the less experienced.

Medicolegal and Psychological Factors

Qualitative feedback revealed a "medicolegal trap." Some experienced readers reported feeling a psychological or legal pressure to follow up on AI-positive findings, even when they suspected the AI was wrong. They feared that overriding an automated "warning" could lead to liability if an aneurysm was later found. Additionally, "satisfaction of search" was noted, where readers might stop their own thorough search once the AI provided a few flags.

Conclusion and Limitations

While AI significantly enhances detection sensitivity and workflow efficiency, it introduces a risk of over-diagnosis and unnecessary invasive procedures due to automation bias.

Limitations: * The small sample size (9 readers, 20 cases) limits generalizability. * The high prevalence of false positives in the study (50%) may not reflect typical clinical practice, potentially influencing reader behavior. * The study did not explicitly measure "omission bias" (missing real aneurysms because the AI did not flag them).

Future implementation of AI in radiology must include training on human-AI interaction and the development of protocols to mitigate bias, ensuring that AI remains a tool for enhancement rather than a source of error.

Summary updated

Abstract

Abstract Background AI systems have the potential to support in detecting cerebral aneurysms. Yet, the role of automation bias (inclination of humans to overly rely on automated decision-making systems) in AI-assisted cerebral aneurysm detection remains unclear. Purpose To determine how automation bias can affect radiologists with varying experience levels when reading time-of-flight magnetic resonance angiography (TOF-MRA) studies with the assistance of an AI system for cerebral aneurysm detection. Methods In this prospective experiment, nine radiologists with varying levels of experience evaluated twenty TOF-MRA exams for the presence of anterior circulation aneurysms, with each arterial segment rated on a 4-point Likert scale, and provided follow-up recommendations. Every case was evaluated twice (with or without assistance by the AI software © mdbrain), with a washout-period of at least four weeks between the two sessions. Ten out of twenty cases included at least one false-positive AI finding. Aneurysm ratings, follow-up recommendations, and reading times were assessed using the Wilcoxon signed-rank test. A thematic analysis was performed to summarize reader feedback and observations. Results False-positive AI results led to significantly higher suspicion of aneurysm findings (p = 0.01). Inexperienced readers further recommended significantly more aggressive follow-up examinations when presented with false-positive AI findings (p = 0.005). Reading times were significantly shorter with AI assistance in inexperienced (164.1 vs 228.2 seconds; p &lt; 0.001), moderately experienced (126.2 vs 156.5 seconds; p &lt; 0.009), and very experienced (117.9 vs 153.5 seconds; p &lt; 0.001) readers alike. Conclusion Our results demonstrate susceptibility of radiology readers to automation bias in detecting cerebral aneurysms in TOF-MRA studies when encountering false-positive AI findings. In inexperienced readers, this behavior further translated into more aggressive follow-up recommendations. AI assistance resulted in significantly shorter reading times across experience levels. While AI systems for cerebral aneurysm detection can provide benefits, challenges in human-AI interaction need to be mitigated to ensure safe and effective adoption.
BibTeX
@misc{kim2024automation,
  author = {Su Hwan Kim and Severin Schramm and Evamaria Olga Riedel and Lena Schmitzer and Enrike Rosenkranz and Olivia Kertels and Jannis Bodden and Karolin J. Paprottka and Dominik Sepp and Martin Renz and Jan S. Kirschke and Thomas Baum and Christian Maegerlein and Tobias Boeckh‐Behrens and Claus Zimmer and Benedikt Wiestler and Dennis M. Hedderich},
  title = {{Automation Bias in AI-Assisted Detection of Cerebral Aneurysms on Time-of-Flight MR-Angiography}},
  year = {2024},
  doi = {10.1101/2024.05.31.24308021},
  howpublished = {medRxiv},
}
Figure