Distinct visual biases affect humans and artificial intelligence in medical imaging diagnoses

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Radiology and diagnostic imaging

Background:

• Artificial intelligence (AI) systems excel at identifying subtle features in diagnostic imaging, potentially surpassing human radiologists.
• These AI capabilities can facilitate earlier disease detection and non-invasive lesion characterization.

Purpose of the Study:

• To explore the differing ways radiologists and AI perceive details in medical images.
• To understand the implications of these perceptual differences for medical decision-making.
• To highlight the risks associated with AI, including bias and lack of causal reasoning.

Main Methods:

• Comparative analysis of AI and radiologist image interpretation processes.
• Exploration of AI's reliance on correlations versus causation.
• Examination of potential demographic and technical biases in AI algorithms.
• Discussion of the interpretability challenges in AI reasoning.

Main Results:

• AI systems detect subtle and higher-order imaging features potentially missed by radiologists.
• AI's reliance on correlations, biases, and uninterpretable reasoning pose risks.
• Significant differences exist in how AI and radiologists learn to interpret medical images.

Conclusions:

• AI offers potential benefits for early disease detection and phenotyping but requires careful validation.
• Understanding the distinct perceptual mechanisms of AI and radiologists is crucial for safe integration.
• Addressing AI biases and ensuring interpretability are key challenges for clinical adoption in medical imaging.