Should we replace radiologists with deep learning? Pigeons, error and trust in medical AI

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Radiology and Oncology

Background:

• Machine learning, particularly deep neural networks, demonstrates high accuracy in identifying malignant cells from radiological images.
• Prominent AI developers suggest replacing human radiologists with AI due to perceived superior accuracy, cost-effectiveness, and predictive power.

Purpose of the Study:

• To critically evaluate the arguments for replacing human radiologists with AI technologies.
• To question the adequacy of current justifications for trusting AI in medical diagnostics.
• To explore the epistemic challenges in establishing AI reliability for medical applications.

Main Methods:

• Philosophical argumentation and analogy (e.g., the pigeon analogy).
• Analysis of the epistemic challenges in verifying the reliability of AI methodologies.
• Examination of the criteria for trusting AI in high-stakes medical settings.

Main Results:

• Arguments based on accuracy, cost, and predictive power are insufficient to justify replacing human radiologists with AI.
• The reliability of AI methodologies like deep neural networks in medical contexts is not yet established and faces fundamental challenges.
• Analogous arguments suggest that even if AI were demonstrably superior, it wouldn't automatically warrant the same level of trust as human experts.

Conclusions:

• Current reasons for deploying AI to replace radiologists are inadequate, even if true.
• Significant epistemic hurdles prevent establishing the reliability required for medical devices.
• There is currently no compelling reason to advocate for replacing radiologists with AI in medical settings.