Multi-reader study benchmarks AI bone age against quantified radiologist variability

Automated BoneXpert bone age was more accurate than single human readers (mean absolute error 4.8 vs 6.5 months) and had fewer >1.8-year errors (0.7% vs 3.4%) in 1,285 pediatric hand radiographs. Reader variability was higher under age 12 (standard deviation 8.7 vs 4.2 months).

• Design: 1,285 left-hand/wrist radiographs from five US academic centers, each read by four radiologists from a pool of 22 across nine institutions.
• Between-institution variability fell from 8.9 to 0.3 months after age adjustment; inter-rater variability was 1.2 months and remained stable after adjustment.
• A deep-learning algorithm performed similarly to human readers (mean absolute error 6.3 months) with 2.4% substantial deviations (>1.8 years), vs 3.4% for single human readings.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.