Radiologists' expertise and LLM model confidence independently affect diagnostic collaboration when using AI for chest imaging.
Key Details
- 1Study involved 10 readers evaluating 100 chest imaging cases (x-ray, CT, MRI, PET) with LLM support.
- 2High-accuracy LLM (text input, GPT-5) achieved 76% accuracy; low-accuracy LLM (vision input, GPT-4o) achieved 27% accuracy.
- 3Model confidence (OR 3.82, p=0.003) and reader expertise (OR 2.06, p<0.001) significantly improved reader-LLM interaction.
- 4High-quality LLM rationales lowered rejection of correct advice (OR 0.79, p=0.005) but raised acceptance of incorrect advice (OR 1.71, p<0.001).
- 5Expertise reduced susceptibility to incorrect suggestions (OR 0.54, p<0.001).
Why It Matters
Understanding the interplay between model confidence, rationale quality, and user expertise is crucial for optimizing LLM-assisted diagnostics in radiology. The study highlights the double-edged nature of LLM 'explanations' and the continuing importance of radiologist training to mitigate overreliance on AI.

Source
AuntMinnie
Related News

•AuntMinnie
Deep Learning PET Biomarkers Outperform Clinical Tools in Prostate Cancer Survival Prediction
Deep learning-derived PET biomarkers more accurately predict metastatic prostate cancer survival than standard clinical risk models.

•AuntMinnie
Multinational AI Model Flags Osteoporosis on Chest X-rays
AI can opportunistically detect osteoporosis from chest x-rays across diverse international populations.

•AuntMinnie
Two-Stage LLM Pipeline Boosts Patient-Friendly Radiology Reports
A two-stage LLM process significantly improves the readability and accuracy of radiology reports for patients.