In this previous post about breast imaging, we briefly touched on the soon-to-be-growing-and-maybe-even-critical problem of automation bias in radiology caused by the growing use of AI.
This study evaluating AI-assisted detection of cerebral aneurysms had similar findings:
Results
False-positive AI results led to significantly higher suspicion of aneurysm findings (p = 0.01). Inexperienced readers further recommended significantly more intense 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 s; p < 0.001), moderately experienced (126.2 vs 156.5 s; p < 0.009), and very experienced (117.9 vs 153.5 s; p < 0.001) readers alike.Conclusion
Our results demonstrate the susceptibility of radiology readers to automation bias in detecting cerebral aneurysms in TOF-MRA studies when encountering false-positive AI findings. 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.
Everyone got faster, but inexperienced readers were fooled by false positives.
This is going to be such a problem.
The reality is that using AI to make us faster is so incredibly ripe for these outcomes. Sure, we could using AI to catch mistakes after an initial independent rad interpretation, and then we could even set up such a system to then use a third party to adjudicate persistent disagreements in a blinded fashion (i.e. a neutral third party radiologist or maybe a different AI agent picks the winner without knowing who they side with)–but the raw economics absolutely point to us using AI as a resident first draft as soon as feasible. It’s going to get messy.
There is an argument that you will have to increasingly be an expert in order to outperform an increasingly competent algorithm. While many current machine mistakes are obvious to experienced radiologists, failures won’t always be comically clear in the future. Assuming we need humans for the long term, training and training quality are critical, and doing so in a way that shields humans from tainting and overreliance on computers will be key.
Yes, pilots use autopilot, but some of those big life-saving stories make the news precisely because pilots also sometimes need to take control.